# Helix (VLA model)

> Source: https://aiwiki.ai/wiki/helix_vla
> Updated: 2026-10-01
> Fact-checked: 2026-09-25
> Categories: AI Models, Embodied AI, Humanoid Robots, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Helix (VLA model)." aiwiki.ai, 1 Oct 2026. https://aiwiki.ai/wiki/helix_vla
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Helix is a [vision-language-action model](https://aiwiki.ai/wiki/vision_language_action_model) (VLA) developed by [Figure AI](https://aiwiki.ai/wiki/figure_ai) that controls humanoid robots by mapping camera images and natural language commands directly to continuous control. It was first announced on February 20, 2025, when Figure described it as the first VLA to output high-rate continuous control of an entire humanoid upper body and the first VLA to run entirely onboard embedded low-power GPUs.[1] Figure says Helix "uses a single set of neural network weights to learn all behaviors", including picking and placing items, using drawers and refrigerators, and cross-robot interaction, without any task-specific fine-tuning.[1] Helix first ran on [Figure 02](https://aiwiki.ai/wiki/figure_02) robots and was later extended to power the [Figure 03](https://aiwiki.ai/wiki/figure_03) platform.

The original Helix uses a dual-system architecture: System 2 (S2), a 7-billion-parameter vision-language model that reasons about the scene and language at 7 to 9 Hz, and System 1 (S1), an 80-million-parameter visuomotor policy that produces fast, reactive control at 200 Hz.[1] At launch Figure claimed several firsts for the model in the vision-language-action model category: the first VLA to output high-rate continuous control of the entire humanoid upper body (a 35-degree-of-freedom action space), the first VLA to operate simultaneously on two robots, and the first VLA to run entirely onboard embedded low-power GPUs.[1] In January 2026, Figure released Helix 02, an extension that added full-body loco-manipulation by introducing a third neural subsystem called System 0.[3] During 2025, Figure extended the same architecture to warehouse package handling, laundry folding, dishwasher loading, and navigation learned from egocentric human video, and in May 2026 the company livestreamed Figure 03 robots running package sorting under Helix 02 for what began as an 8-hour shift and ended at 200 hours.[18][19][20][25][28][34] On September 17, 2026, Figure introduced Helix 2.5, a model pretrained from scratch on its Index human-video dataset, and reported that it completed 237 of 420 whole-body household trials (56%) in 30 Bay Area homes where no training data had been collected.[28]

## Overview

Helix is a generalist control model for [humanoid robots](https://aiwiki.ai/wiki/humanoid_robot): a single neural network that takes monocular camera images, robot state, and a natural language instruction and outputs continuous control of the robot's wrists, torso, head, and individual fingers. Figure states that "Helix is the first VLA to output high-rate continuous control of the entire humanoid upper body, including wrists, torso, head, and individual fingers," and the first VLA "that runs entirely onboard embedded low-power-consumption GPUs, making it immediately ready for commercial deployment."[1] Rather than a library of hand-coded skills, Helix learns picking, placing, drawer and refrigerator use, and two-robot handovers with one set of weights and no per-task fine-tuning.[1]

The model has gone through three publicly named generations:

| Version | Announced | Main change | Robot |
|---|---|---|---|
| Helix | February 20, 2025 | S2 (7B VLM, 7-9 Hz) plus S1 (80M visuomotor policy, 200 Hz) for upper-body control[1] | Figure 02 |
| Helix 02 | January 27, 2026 | Adds System 0 (10M parameters, 1 kHz) for learned whole-body control; S1 outputs full-body joint targets[3] | Figure 03 |
| Helix 2.5 | September 17, 2026 | Pretrained from random initialization on Index human-behavior data, then adapted to whole-body household tasks; evaluated zero-shot in 30 homes[28] | Figure 03 |

## Background

Figure AI is a humanoid robot company led by its founder and CEO [Brett Adcock](https://aiwiki.ai/wiki/brett_adcock). Its first-generation robot, Figure 01, took its first steps in May 2023.[40] Its second, [Figure 02](https://aiwiki.ai/wiki/figure_02), was built for real-world deployment; in August 2024 BMW Group said Figure 02 had completed a trial run lasting several weeks at BMW Group Plant Spartanburg in South Carolina, inserting sheet metal parts into fixtures.[6][40]

Figure's own framing of the problem Helix was meant to solve is that teaching robots even a single new behavior required "either hours of PhD-level expert manual programming or thousands of demonstrations."[1] Before Helix 02, Figure's robots also relied on hand-engineered C++ code for lower-body control and balance.[3][9] [RT-2](https://aiwiki.ai/wiki/rt_2) and similar work had shown that a vision-language model pretrained on web data could be fine-tuned to output robot actions,[15] and Figure's stated goal with Helix was to "translate the rich semantic knowledge captured in Vision Language Models (VLMs) directly into robot actions."[1]

Helix was tied to Figure's decision to bring its robot intelligence fully in-house. On February 4, 2025, Adcock announced on X that he had decided to leave Figure's collaboration agreement with [OpenAI](https://aiwiki.ai/wiki/openai), writing: "Figure made a major breakthrough on fully end-to-end robot AI, built entirely in-house. We're excited to show you in the next 30 days something no one has ever seen on a humanoid."[12] He told TechCrunch: "We found that to solve embodied AI at scale in the real world, you have to vertically integrate robot AI. We can't outsource AI for the same reason we can't outsource our hardware."[37] On February 20, 2025, Figure published a detailed technical post and a demonstration video in which two Figure 02 robots put away groceries neither robot had seen before.[1][10]

On May 29, 2025, Adcock announced what he called the "largest re-org in Figure's history," consolidating three separate teams into a unified AI group also named Helix.[38]

Capital followed. On September 16, 2025, Figure announced that it had exceeded $1 billion in committed Series C funding at a $39 billion post-money valuation, in a round led by Parkway Venture Capital with significant investment from Brookfield Asset Management, [NVIDIA](https://aiwiki.ai/wiki/nvidia), Macquarie Capital, Intel Capital, Align Ventures, Tamarack Global, LG Technology Ventures, [Salesforce](https://aiwiki.ai/wiki/salesforce), T-Mobile Ventures, and Qualcomm Ventures. The company said the funds would scale production at BotQ and real-world deployments, build GPU infrastructure for Helix training and simulation, and finance data collection of human video and multimodal sensory inputs.[21]

## Architecture

Helix uses a dual-system design that Figure describes as a "System 1, System 2" VLA, in which S2 can "think slow" about high-level goals while S1 can "think fast" to execute and adjust actions.[1] The two subsystems operate at very different timescales and communicate through a latent vector in shared memory.[1]

### System 2: the semantic reasoner

System 2 is built on a 7-billion-parameter open-source, open-weight vision-language model (VLM) pretrained on internet-scale data. It processes monocular robot images and robot state information (wrist pose and finger positions), projected into the vision-language embedding space, together with natural language commands. It runs at 7 to 9 Hz.[1]

S2's job is scene understanding and language comprehension: it distills all semantic, task-relevant information into a single continuous latent vector, which is passed to S1 to condition its low-level actions.[1] Figure attributes Helix's ability to generalize across objects and contexts to this internet-pretrained backbone.[1]

### System 1: the visuomotor controller

System 1 is an 80-million-parameter cross-attention encoder-decoder transformer. It uses a fully convolutional, multi-scale vision backbone initialized from pretraining done entirely in simulation. S1 receives the same image and state inputs as S2 but processes them at a higher frequency; the latent vector from S2 is projected into S1's token space and concatenated with S1's visual features.[1]

S1 outputs full upper-body humanoid control at 200 Hz: desired wrist poses, finger flexion and abduction control, and torso and head orientation targets. Figure also appends a synthetic "percentage task completion" action, which lets Helix predict its own termination condition and makes it easier to sequence multiple learned behaviors.[1] Figure reports that Helix coordinates a 35-DoF action space at 200 Hz.[1] Figure says directly outputting continuous control avoids the action tokenization schemes used by earlier VLAs, which it says face scaling challenges in high-dimensional humanoid control.[1]

During training, Figure adds a temporal offset between S1 and S2 inputs, calibrated to match the gap between the two systems' deployed inference latency, so that real-time control conditions during deployment are reflected in training.[1]

S2 runs as an asynchronous background process, continuously updating the shared-memory latent vector. S1 runs as a separate real-time process that maintains the 200 Hz control loop, using the latest observation and the most recent S2 latent.[1] At 200 Hz, each S1 control cycle is 5 milliseconds.

### Onboard inference

Each Figure robot used for Helix is equipped with dual low-power embedded GPUs, and the inference pipeline is split so that S2 and S1 each run on a dedicated GPU, bridged by the shared-memory latent vector.[1] Third-party spec listings describe Figure 02's onboard compute as dual NVIDIA RTX GPU-based modules,[27] and in September 2026 NVIDIA CEO Jensen Huang described Figure's models as being deployed "on NVIDIA GPUs in Figure's robots."[33] Figure says this deployment strategy lets it run Helix "as fast as our fastest single task imitation learning policies," and that running entirely onboard makes the model ready for commercial deployment.[1]

## Training

Figure trained the original Helix on a high-quality, multi-robot, multi-operator dataset of diverse teleoperated behaviors, about 500 hours in total.[1] To create language-conditioned training pairs, the company used hindsight instruction labeling: an auto-labeling VLM processed segmented video clips from the onboard cameras and was prompted with "What instruction would you have given the robot to get the action seen in this video?"[1]

Figure described the training set as "a small fraction of the size of previously collected VLA datasets (<5%)," achieved without multi-robot-embodiment data collection or multiple training stages.[1]

Helix is trained fully end-to-end with a standard regression loss, mapping raw pixels and text commands to continuous actions. Gradients are backpropagated from S1 into S2 through the latent communication vector, so both components are optimized jointly. There are no task-specific fine-tuning stages or separate action heads for different behaviors.[1]

Figure says all items handled during training were excluded from evaluations to prevent contamination.[1]

### Project Go-Big: human video pretraining (September 2025)

On September 18, 2025, Figure announced Project Go-Big, which it described as an effort to build the world's largest and most diverse humanoid pretraining dataset, centered on egocentric (first-person) human video rather than robot teleoperation.[20] The initiative is supported by a partnership with Brookfield Asset Management, which Figure said has more than 100,000 residential units, 500 million square feet of commercial office space, and 160 million square feet of logistics space.[20] As a first result, Figure reported that Helix learned to navigate cluttered real-world spaces from natural language commands, mapping images and language directly to low-level SE(2) velocity commands, after training on 100% egocentric human video collected in Brookfield homes. The company said that, to its knowledge, this was the first time a humanoid robot had learned end-to-end navigation using only human video, with one Helix network outputting both dexterous manipulation and navigation.[20]

## Capabilities

### Object generalization

The most widely discussed capability of Helix at launch was its ability to handle objects the model had never seen during training. Figure reported that robots running Helix can pick up virtually any small household object with a simple "Pick up the [X]" command, and that in systematic testing they handled thousands of novel items in clutter, from glassware and toys to tools and clothing, without prior demonstrations.[1] In one example, when prompted to "Pick up the desert item," Helix identified a toy cactus as matching the concept, selected the closer hand, and grasped it.[1]

Figure attributes this generalization to S2's internet-pretrained VLM backbone.[1]

### Multi-robot collaboration

Figure says Helix is the first VLA to operate simultaneously on two robots.[1] In the February 2025 demonstration, two Figure 02 robots used identical Helix weights to put away groceries neither had encountered in training, handing items between them.[1][10] Figure said this required no robot-specific training or explicit role assignments; the robots were coordinated through natural language prompts such as "Hand the bag of cookies to the robot on your right" and "Receive the bag of cookies from the robot on your left and place it in the open drawer."[1]

### Whole upper-body control

Helix integrates the 35 upper-body degrees of freedom into a single action output, controlling everything from individual finger movements to end-effector trajectories, head gaze, and torso posture.[1] Figure notes that head and torso control are hard because moving them changes both what the robot can reach and what it can see, creating feedback loops that have historically caused instability; in its demonstration the robot tracks its hands with its head while adjusting its torso for reach.[1] Figure says that, to its knowledge, no prior VLA had shown this degree of real-time coordination while generalizing across tasks and objects.[1]

### Onboard deployment

Running S2 and S1 entirely on embedded GPU hardware, without cloud inference, was a stated commercial priority at launch. Figure framed this as the model being "immediately ready for commercial deployment."[1] During the May 2026 package-sorting livestream, Humanoids Daily reported that Helix-02 runs entirely locally on an onboard computer housed in the robot's torso, and that Adcock told Bloomberg Technology there was "absolutely no teleoperation" in the demonstration.[41]

### Deformable object manipulation: laundry folding (August 2025)

On August 12, 2025, Figure reported that Helix had learned to fold laundry, which the company described as the first instance of a humanoid robot with multi-fingered hands folding laundry fully autonomously using an end-to-end neural network.[19] Figure stated that the same Helix architecture used for logistics was applied with no modifications to the model or training hyperparameters; only the dataset changed. Demonstrated behaviors included picking towels from a mixed pile, adapting the folding strategy to each towel's starting configuration, recovering from multi-pick errors by returning extra items, and fine motions such as tracing an edge with a thumb, pinching corners, or unraveling tangled towels before completing folds.[19] Figure also said Helix learned to maintain eye contact, direct its gaze, and use learned hand gestures while engaging with people.[19]

### Dishwasher loading (September 2025)

On September 3, 2025, Figure reported that the same Helix model could load a dishwasher with "no new algorithms, no special-case engineering, just new data." Demonstrated behaviors included singulating stacked plates, picking up a glass with one hand and reorienting it before placing it with the other, adjusting to messy starting configurations, and recovering from misgrasps or collisions.[34]

## Helix logistics variant (February 26, 2025)

Six days after the initial announcement, Figure published a follow-on technical report on logistics package manipulation and triaging, focused on general improvements to System 1.[2]

First, Figure replaced S1's monocular input with an implicit stereo vision backbone and a multi-scale feature extraction network. Features from both cameras are merged before tokenization so the number of visual tokens stays constant. Figure reported that the stereo model achieved a 60% increase in throughput over non-stereo baselines and generalized to flat envelopes that the system was never trained on.[2]

Second, Figure trained a learned visual proprioception model that estimates the 6D poses of the end effectors from each robot's onboard cameras. This online "self-calibration" let a policy trained on a single robot's data transfer to additional robots with comparable manipulation performance, reducing the need for per-robot recalibration.[2]

On data, Figure filtered out slower, missed, or failed human demonstrations, but deliberately kept demonstrations that included corrective behavior when the failure was due to environmental stochasticity rather than operator error. It also said working closely with teleoperators to refine and standardize manipulation strategies produced significant improvements.[2] Figure reported that just 8 hours of well-curated demonstration data could yield a dexterous and flexible policy, and that a model trained on curated, high-quality demonstrations achieved 40% better throughput despite being trained with one-third less data.[2]

A "Sport Mode" feature speeds up execution at test time by linearly resampling each S1 action chunk to a shorter trajectory and executing it at the original 200 Hz rate; for example, a 20% speedup requires no change to training.[2] Figure reported that this worked well up to a 50% speedup, at which point the policy handled packages faster than the expert demonstrations it was trained on, and that beyond 50% throughput dropped substantially as motions became too imprecise.[2]

### Scaling update (June 2025)

A second logistics post, on June 7, 2025, reported gains from both data scaling and architecture changes. Handling time fell to 4.05 seconds per package, down from about 5.0 seconds (about 20% faster), and Figure said the policy now handled deformable poly bags and flat envelopes as reliably as rigid boxes.[18] Shipping labels were correctly oriented for scanning about 95% of the time, up from about 70%.[18] The update added three elements to System 1: a vision memory module that composes features from recent frames, a history of recent proprioceptive states (hand, torso, and head positions), and force feedback in the state input as a proxy for touch.[18] In a data-scaling study with models trained on about 10, 20, 40, and 60 hours of demonstrations, going from 10 to 60 hours cut average handling time from about 6.84 seconds to 4.31 seconds (a 58% throughput increase) and raised barcode success from 88.2% to 94.4%. Figure said these returns suggested the model was still in a low-data regime.[18] Enlarging the transformer decoder head by 50% brought the average to 4.05 seconds while holding accuracy above 92%.[18]

## Helix 02 (January 2026)

On January 27, 2026, Figure released Helix 02, demonstrated on the [Figure 03](https://aiwiki.ai/wiki/figure_03) platform. It introduced a third subsystem called System 0 (S0) that extended Helix from upper-body manipulation to full-body loco-manipulation.[3] Figure called Helix 02 its "most capable humanoid model yet."[3]

### System 0

System 0 is a 10-million-parameter neural network that takes full-body joint state and base motion as input and outputs joint-level actuator commands at 1 kHz, handling balance, contact, and coordination across the entire body.[3] Figure describes it as a learned whole-body controller that replaces 109,504 lines of hand-engineered C++.[3]

S0 was trained on over 1,000 hours of joint-level retargeted human motion data, entirely in simulation across more than 200,000 parallel environments with extensive domain randomization, using sim-to-real [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning).[3] In an interview reported by Humanoids Daily, Adcock described deleting the final 109,504 lines of hand-engineered C++ from the robots as a "Software 2.0" milestone.[9]

The three-tier hierarchy in Helix 02 is:

- S2: interprets scenes, understands language, and sequences behaviors
- S1 runs at 200 Hz: translates perception into full-body joint targets
- S0 runs at 1 kHz: handles balance, contact, and whole-body coordination[3]

In Helix 02, S1 connects to all sensors (head cameras, palm cameras, fingertip tactile sensors, and full-body proprioception) and outputs joint-level control of the entire robot, including legs, torso, head, arms, wrists, and individual fingers. S1 remains a transformer conditioned on S2 latents.[3]

### Full-body manipulation demonstrations

The flagship demonstration for Helix 02 showed a Figure 03 robot unloading and reloading a dishwasher across a full-sized kitchen: a four-minute, end-to-end autonomous task with no resets and no human intervention.[3] In the video the robot walks to the dishwasher, unloads dishes, navigates across the room, stacks items in cabinets, and then loads and starts the dishwasher; Figure counted 61 loco-manipulation actions.[3] The robot also closed a drawer with its hip and lifted the dishwasher door with its foot when its hands were occupied.[3] Figure said it believed this was the longest-horizon, most complex task completed autonomously by a humanoid robot to date.[3]

Other demonstrated tasks included unscrewing a bottle cap, picking a pill out of a medicine box, pushing exactly 5 ml from a syringe, and picking small metal pieces from a cluttered box.[3] Figure said these tasks depend on the tactile sensing and palm cameras of Figure 03, and that this was the first time it had demonstrated neural network policies that depend on those modalities.[3]

### Sensing upgrades in Figure 03

Figure says Figure 03 was designed for Helix. Its camera architecture delivers twice the frame rate, one-quarter the latency, and a 60% wider field of view per camera compared with its predecessor.[4] Palm-mounted cameras in each hand provide close-range visual feedback during grasping, including when the main cameras are occluded.[4] Figure-developed fingertip tactile sensors detect forces as small as 3 grams, which Figure compares to the weight of a paperclip.[4] These sensors are part of S1's inputs in Helix 02.[3]

### Living room and bedroom demonstrations (March and May 2026)

On March 9, 2026, Figure showed Helix 02 tidying a living room: spraying and wiping a surface with a towel, whipping the towel over its shoulder to free its hands, holding a bin with both hands while scooping blocks into it, tucking a container under one arm, tossing a pillow onto a couch, reorienting a remote in-hand to press the button that turns off a TV, and side-stepping through the gap between a coffee table and couch. Figure said these behaviors came from new data with no new algorithms.[30]

On May 8, 2026, Figure showed two Helix 02 robots resetting a bedroom in under two minutes: opening doors, hanging clothes, putting away headphones, closing a book, taking out trash via a foot-pedal bin, pushing a chair under a desk, and making a bed together. Figure said both robots ran a single learned VLA policy with no shared planner, message passing, or central coordinator, each inferring its partner's intent from its own cameras, and called it, to its knowledge, the first demonstration of a single learned neural network performing multi-humanoid collaborative loco-manipulation directly from pixels to actions.[31]

### Perception-conditioned System 0 (April 2026)

In an April 29, 2026 production update, Figure said S0 had previously reasoned only about the robot's own body, so stairs, ramps, and uneven terrain required hand-tuned mode switches and operator intervention. S0 is now conditioned on camera perception: head-camera RGB images pass through Figure's stereo model into a 3D representation that is fed to the policy alongside proprioception. Figure said the policy was trained end-to-end with reinforcement learning in simulation across thousands of randomized terrains and transferred zero-shot to real-world stairs.[24]

### Endurance demonstrations (May 2026)

In May 2026, Figure used Helix 02 to demonstrate sustained autonomous work on small-package sorting, in which each Figure 03 detects a package's barcode, picks it up, and places it barcode-face-down on a conveyor.[25][26] On May 13, Adcock posted: "Watch a team of humanoid robots running a full 8-hr shift at human performance levels. This is fully autonomous running Helix-02."[25] Figure kept the livestream running after the first shift. Adcock wrote that the robots were "over 24 hours of continuous autonomous operation without a failure," and Interesting Engineering reported that three robots, nicknamed Bob, Frank, and Gary, had sorted more than 28,000 packages by then.[26] Adcock said humans average around 3 seconds per package and that Figure 03 was "now around human parity," stated that there was no teleoperation, and said Helix triggers an automatic reset if a robot gets stuck or the policy goes out of distribution, while a robot with a software or hardware issue leaves for maintenance and another takes over.[26]

The run eventually lasted 200 hours: Figure's September 2026 Helix 2.5 announcement refers to Helix 02 "running a logistics task autonomously for 200 hours."[28] Humanoids Daily reported that the livestream ended at the 200-hour mark with 249,560 packages on its dashboard, quoted Adcock as saying the run went "200 hours without a failure," and noted that viewers saw occasional package-level errors, such as packages falling off the belt.[36]

## Index and Helix 2.5 (August-September 2026)

### Index

On August 25, 2026, Figure publicly launched [Index](https://aiwiki.ai/wiki/figure_index), a data-collection app that had operated in stealth for about four months. People Figure calls Creators record themselves doing real tasks at home or work; Figure described Index as "the largest useful robot training dataset in the world."[29] At launch Figure reported 264,000 downloads across 108 countries, more than 44,000 weekly active users, more than 16 million uploaded videos, 30 minutes of video uploaded every second, and $15 million paid to Creators, and said it was committed to spending over $1 billion over the next 12 months on data and compute.[29] Per 1,000 hours collected, Figure said Index contains 373 unique tasks, 1,146 unique manipulated objects, and 116 unique environments. Uploads pass through a five-stage pipeline of filtering, fraud review, deduplication, rebalancing, and annotation.[29] Figure said it had tried buying data first, but vendors could not meet Helix's throughput, diversity, or quality bar.[29]

On September 3, 2026, Figure announced a partnership with [Nscale](https://aiwiki.ai/wiki/nscale) to deploy up to 100,000 GPUs on the [NVIDIA Vera Rubin](https://aiwiki.ai/wiki/nvidia_vera_rubin) platform, with initial deployment targeted from the second half of 2027. Figure called it an initial commitment of $3.5 billion of compute, with intent to scale to over $6 billion, to train Helix.[33]

### Helix 2.5

On September 17, 2026, Figure introduced Helix 2.5, which it called "the most advanced neural network Figure has built."[28] Unlike Helix 02, which started from a pretrained vision-language model, Helix 2.5 was pretrained from random initialization entirely on Index. From that single foundation model Figure produced three whole-body behaviors (tidying a living room, folding towels, and making a bed) by fine-tuning on task-specification data collected elsewhere, then took the robots into 30 Bay Area homes with zero data collected in any of them.[28] Adcock said Figure rented the homes for the experiment.[35]

Figure defines "zero-shot" as referring to the evaluation environments and manipulated objects: no data was collected in any evaluation home, no evaluation toy, towel, or bedding appeared in task-specification data, and the robot used each home's own couches, beds, and folding surfaces. Each task used a single fixed checkpoint across all 30 homes, and no evaluation rollouts were used for checkpoint selection. Evaluation objects were set aside in advance and checked by an AI model and human review.[28]

Success criteria were fixed before evaluation and gave no partial credit:[28]

- Living room tidy: all 13-15 toys scattered in the scene are picked and placed in the basket (1-minute timeout per toy).
- Towel folding: all towels are folded and placed in the basket (3-minute timeout per towel).
- Bed making: both pillows and the comforter corners are placed at the top of the bed, with the comforter pulled smooth.
- Any rollout that needed a human intervention for safety was counted as a failure.

To measure the contribution of Index, Figure trained two policies on identical task-specification data, with architecture, optimization, hyperparameters, and evaluation held fixed; one started from random weights and the other from the Index-pretrained Helix 2.5 model.[28] Figure's results chart reports:

| Task (chart label) | Trained without Index | Trained with Index |
|---|---|---|
| Towel folding (all 4 towels) | 12/140 (9%) | 87/140 (62%) |
| Living room tidy (every toy) | 7/140 (5%) | 56/140 (40%) |
| Bed making (all 3 items, graded B or better) | 16/140 (11%) | 94/140 (67%) |
| All three tasks, pooled | 35/420 (8%) | 237/420 (56%) |

Source: Figure, "Zero-Shot Success in 30 Homes" chart; success is measured per rollout.[28]

Figure's text gives the from-scratch baseline as 9%, while its chart shows 35 of 420 trials (about 8%); Humanoids Daily pointed out this small discrepancy.[28][35] Figure said no single evaluation task makes up more than 1.90% of the Index pretraining dataset.[28]

Figure made several further claims for Helix 2.5:[28]

- Data efficiency: compared with a Helix 02 policy for the same task that was trained on data collected in its evaluation environment, Helix 2.5 matched that policy's success rate with half as much adaptation data, while running zero-shot in 30 unseen homes.
- Self-correction: the robot steps back to reposition, changes stance, or walks around a bed to fix a fold, which Figure attributes to Index pretraining.
- A scaling law: four models trained on nested subsets of Index spanning an 8x increase in pretraining data, with model size and downstream training fixed, showed held-out robot-action prediction loss falling predictably with each doubling. Using only the smaller runs, Figure said it forecast the largest run's test loss to four decimal places, with forecasting error of 0.54% of the variation across the 8x range. Figure called this, to its knowledge, the first human-to-robot transfer scaling law measured on a humanoid.

Figure said that, to its knowledge, Helix 2.5 is the first demonstration of zero-shot whole-body generalization at this scope on a humanoid, while adding that "the point is not that general humanoid robotics is solved."[28] The post also said Index was by then generating roughly 35 minutes of new human experience every second and that Figure had committed $3.5 billion of compute to training Helix.[28]

Humanoids Daily described the result as a company-run evaluation that supports transfer to unfamiliar settings while leaving a substantial reliability gap: roughly 44% of trials were not completed under Figure's criteria. It reported that Adcock posted nearly four hours of extended footage from the homes, which is not a complete record of the 420 trials, and that Sunday Robotics' Tony Zhao criticized the framing as "failing half the time," arguing that "Doing useful work = generalization + reliability."[35] The outlet noted that Sunday's own garment-folding success rate is not directly comparable, because it counts individual garments whereas Figure's trials require the whole task to be completed.[35]

## Robot platforms

### Figure 02

[Figure 02](https://aiwiki.ai/wiki/figure_02) was a bipedal humanoid about 170 cm tall and 70 kg, with 16-degree-of-freedom hands, per BMW Group's published summary.[27] It was the platform for Helix at launch and for the BMW deployment.[1][5] Figure began retiring Figure 02 after the release of Figure 03, and on September 30, 2026 said most of the fleet had been melted down at a foundry in Imatra, Finland, with only a few units left in storage at its headquarters.[5][46]

### Figure 03

[Figure 03](https://aiwiki.ai/wiki/figure_03) was announced on October 9, 2025,[4] with Figure saying it was designed "for Helix, the home, and the world at scale." Figure lists it at 5 feet 8 inches tall and 61 kilograms,[39] and says it has 9% less mass and significantly less volume than Figure 02.[4] Figure 03 supports wireless inductive charging at 2 kW through coils in its feet, so the robot can step onto a stand to recharge. Its actuators run at twice the speed of Figure 02's with improved torque density, and the body uses multi-density foam and soft textiles rather than hard machined parts.[4] Figure 03's 10 Gbps mmWave data offload lets the fleet upload terabytes of data for continuous learning.[4]

Figure said BotQ's first-generation line would initially be able to produce up to 12,000 humanoid robots per year, with a goal of producing 100,000 robots over four years.[4] Humanoids Daily reported in April 2026 that Adcock had said on the Shawn Ryan Show that BotQ was producing a robot every 90 minutes, and that Figure aims to scale to 50,000 units a year.[23]

TIME named Figure 03 one of its Best Inventions of 2025, noting that at launch the robot could do some domestic tasks such as folding clothes and loading a dishwasher, but not without help, and that Adcock aimed to have Figure 03 in select homes the following year.[16] In an April 29, 2026 update, Figure said BotQ had gone from one Figure 03 per day to one per hour, a 24x throughput improvement in under 120 days, with more than 350 Figure 03 units delivered, end-of-line first-pass yield above 80%, and custom manufacturing software running across more than 150 networked workstations.[24] Figure frames the growing fleet as a data source for Helix, and said its over-the-air update system can deploy new behaviors and upgrades to the entire fleet simultaneously.[24]

## BMW factory deployment

The most extensively documented real-world deployment of a Figure robot was the Figure 02 program at BMW Group Plant Spartanburg in South Carolina. Neither Figure's report on the program nor BMW's statements about it name the AI model that controlled the robots.[5][7] BMW said in August 2024 that Figure 02 had completed a trial run lasting several weeks at the plant.[6] In its November 19, 2025 report on the 11-month deployment, Figure said it delivered robots to the plant within 6 months of bringing up Figure 02 and launched full deployment on an active assembly line within 10 months, running every working day with 10-hour shifts Monday through Friday.[5]

Figure 02 robots performed sheet-metal loading: picking parts from racks or bins and placing them on a welding fixture within a 5-millimeter tolerance in 2 seconds.[5] Figure reported the following totals:[5]

- 90,000+ parts loaded
- Contributed to the production of 30,000+ BMW X3 vehicles
- 1,250+ hours of runtime
- An estimated 1.2+ million robot steps (200+ miles)

BMW's own February 2026 statement gives consistent figures: within ten months, Figure 02 supported the production of more than 30,000 BMW X3s, working ten-hour shifts Monday to Friday, moving more than 90,000 components and covering about 1.2 million steps in around 1,250 operating hours.[7]

Figure defined three KPIs: an 84-second cycle time (37 seconds for loading), greater than 99% placement accuracy per shift, and zero human interventions per shift.[5]

Figure said Figure 02 recorded minimal hardware failures across more than 1,250 operating hours, and that the forearm was the top hardware failure point, which it attributed to tight packaging, three degrees of freedom, and thermal constraints. For Figure 03, the company re-architected the wrist electronics to eliminate the distribution board and dynamic cabling, with each wrist motor controller communicating directly with the main computer.[5]

In February 2026, BMW announced a separate humanoid pilot with a different robot maker, Hexagon Robotics. An initial test deployment of Hexagon's [AEON](https://aiwiki.ai/wiki/hexagon_aeon) robot took place at BMW Group Plant Leipzig in Germany in December 2025, with a further test deployment planned from April 2026 ahead of a pilot phase starting in summer 2026, the first deployment of humanoid robots in BMW's production in Germany.[7] BMW also said it and Figure were evaluating additional use cases for Figure 03.[7]

### Figure 03 at BMW (June 2026)

On June 30, 2026, Figure said Figure 03 had arrived in Hall 52, an assembly and logistics hall at Spartanburg, for a parts-sequencing use case. Figure said Helix 02 coordinates the robot's hands, arms, torso, and feet so it can pick and place thin-walled parts while stepping and repositioning, and pull a heavy cart on caster wheels.[32]

### Other commercial deployments

On May 26, 2026, Figure announced a commercial agreement with Catalyst Brands to deploy Figure humanoids in its distribution and logistics network, starting at its Reno, Nevada Distribution Logistics Center.[45]

## Comparison with other VLA models

Helix belongs to a generation of VLA models for robot control that emerged between 2023 and 2026. The models differ substantially in architecture, parameter count, control frequency, and target robot platforms.

| Model | Developer | Release | Parameters | Control rate | Architecture type | Target platform |
|---|---|---|---|---|---|---|
| [RT-2](https://aiwiki.ai/wiki/rt_2) | Google DeepMind | July 2023 | 5B and 55B (PaLI-X); 12B (PaLM-E) | 1-3 Hz (55B); about 5 Hz (5B) | Single end-to-end VLA | 7-DoF mobile manipulator |
| Helix (S1+S2) | Figure AI | Feb 2025 | 7B (S2) + 80M (S1) | 200 Hz | Dual-system (VLM + visuomotor) | Humanoid upper body |
| [Isaac GR00T N1](https://aiwiki.ai/wiki/isaac_gr00t) | NVIDIA | Mar 2025 | 2.2B (GR00T-N1-2B) | 120 Hz (VLM at 10 Hz) | Dual-system (VLM + diffusion transformer) | Humanoid (cross-embodiment) |
| [pi0](https://aiwiki.ai/wiki/pi0) | Physical Intelligence | Oct 2024 | 3.3B | Up to 50 Hz | VLM + action expert (flow matching) | Single-arm, dual-arm, and mobile manipulators |
| GR00T N1.7 | NVIDIA | Apr 2026 (early access) | 3B | Not stated | Dual-system (Cosmos-Reason2-2B VLM + diffusion transformer) | Humanoid |

RT-2 demonstrated that a VLM pretrained on internet data could be co-fine-tuned on robot demonstrations to produce a general-purpose robot policy.[15] Its largest model, RT-2-PaLI-X-55B, ran at 1-3 Hz and its 5B version at about 5 Hz, served from a multi-TPU cloud service and queried over the network rather than run onboard; the robot was a 7-DoF mobile manipulator.[15]

[pi0](https://aiwiki.ai/wiki/pi0) from Physical Intelligence uses a 3.3-billion-parameter architecture that combines a PaliGemma VLM backbone with a flow-matching action expert, producing action chunks at up to 50 Hz.[14] It was pretrained on over 10,000 hours of robot data plus the open-source OXE dataset and evaluated on single-arm, dual-arm, and mobile manipulators.[14] Physical Intelligence publishes pi0 and the later [pi0.5](https://aiwiki.ai/wiki/pi_0_5) as open-source models in its openpi repository, whereas Helix is proprietary to Figure's robots.[42][4]

[Isaac GR00T N1](https://aiwiki.ai/wiki/isaac_gr00t) from NVIDIA uses a dual-system design that parallels Helix's: a VLM module running at 10 Hz coupled with a diffusion transformer that generates closed-loop motor actions at 120 Hz. Its publicly released GR00T-N1-2B model has 2.2 billion parameters, 1.34 billion of them in the VLM.[44] It was trained on a mixture of real-robot trajectories, human videos, and synthetic data generated with physics simulation and neural video models, and released as an open foundation model.[44][13] NVIDIA's GR00T N1.7, released in early access in April 2026, is a 3B-parameter, commercially licensed model pretrained on 20,854 hours of human egocentric video; NVIDIA described its results as "the first-ever scaling law for robot dexterity."[43] Figure's September 2026 Helix 2.5 post likewise described a human-to-robot transfer scaling law, measured on action-prediction loss.[28]

Helix's S1 runs at 200 Hz, compared with 120 Hz for GR00T N1's action module, up to 50 Hz for pi0, and 1-5 Hz for RT-2, and it controls a 35-DoF upper-body action space.[1][14][15][44] Figure says this high-rate continuous output is what lets Helix handle high-dimensional humanoid control without action tokenization.[1]

The training data profiles also differ. The original Helix policy was trained on about 500 hours of real teleoperated demonstrations, with only S1's vision backbone initialized from simulation pretraining.[1] Helix 02 added S0, trained entirely in simulation on retargeted human motion data.[3] Helix 2.5 moved to pretraining from scratch on Index human video,[28] a direction also taken by Project Go-Big and by NVIDIA's GR00T N1.7.[20][43]

## Limitations and criticism

Several limitations in Helix's capabilities and deployment context appear in Figure's own technical disclosures and in outside coverage.

The original Helix controlled only the upper body; walking and balance were not part of the Helix network until Helix 02 added the learned S0 controller.[1][3] Until April 2026, S0 itself could not see the terrain in front of the robot, so stairs, ramps, and uneven terrain required hand-tuned mode switches and operator intervention.[24]

The BMW deployment surfaced hardware issues, above all the Figure 02 forearm, which Figure addressed by re-architecting the wrist electronics for Figure 03.[5]

Helix still depends on task-specific data to specify each behavior. Figure's logistics work showed that curated data mattered more than raw quantity for a single use case,[2] and even Helix 2.5 requires task-specification fine-tuning data for each behavior.[28] Helix 2.5's best pooled result completed 56% of whole-task trials in unseen homes, which Humanoids Daily noted leaves roughly 44% of trials unfinished,[35] and Figure's scaling-law result measures action-prediction loss rather than task success.[28]

In November 2025, Figure's former principal robotic safety engineer, Robert Gruendel, filed a lawsuit in federal court in the Northern District of California alleging that he was wrongfully terminated in September 2025 after warning executives that the robots "were powerful enough to fracture a human skull." The complaint also alleged that a robot had "carved a ¼-inch gash into a steel refrigerator door during a malfunction." A Figure spokesperson said Gruendel was "terminated for poor performance" and that his "allegations are falsehoods that Figure will thoroughly discredit in court." Gruendel's attorney said the case "may be among the first whistleblower cases related to the safety of humanoid robots."[22]

For home use, Adcock said in a February 2026 interview that he still "babysits" the robot around his own children, adding: "Until I feel safe enough to have it there with free reign, it's not ready for everyone."[9] As of September 2026, Figure had shown Helix 2.5 working in 30 rented homes; neither its Helix 2.5 announcement nor Humanoids Daily's coverage gave a date for consumer availability.[28][35]

## See also

- [Figure 02](https://aiwiki.ai/wiki/figure_02)
- [Figure 03](https://aiwiki.ai/wiki/figure_03)
- [Index (Figure AI)](https://aiwiki.ai/wiki/figure_index)
- [Isaac GR00T](https://aiwiki.ai/wiki/isaac_gr00t)
- [pi0](https://aiwiki.ai/wiki/pi0)
- [RT-2](https://aiwiki.ai/wiki/rt_2)
- [Vision-language-action model](https://aiwiki.ai/wiki/vision_language_action_model)
- [Embodied AI](https://aiwiki.ai/wiki/embodied_ai)
- [Figure AI](https://aiwiki.ai/wiki/figure_ai)

## References

1. Figure AI. "Helix: A Vision-Language-Action Model for Generalist Humanoid Control." figure.ai, February 20, 2025. https://www.figure.ai/news/helix
2. Figure AI. "Helix Accelerating Real-World Logistics." figure.ai, February 26, 2025. https://www.figure.ai/news/helix-logistics
3. Figure AI. "Introducing Helix 02: Full-Body Autonomy." figure.ai, January 27, 2026. https://www.figure.ai/news/helix-02
4. Figure AI. "Introducing Figure 03." figure.ai, October 9, 2025. https://www.figure.ai/news/introducing-figure-03
5. Figure AI. "F.02 Contributed to the Production of 30,000 Cars at BMW." figure.ai, November 19, 2025. https://www.figure.ai/news/production-at-bmw
6. BMW Group. "Successful test of humanoid robots at BMW Group Plant Spartanburg." press.bmwgroup.com, August 6, 2024. https://www.press.bmwgroup.com/global/article/detail/T0444265EN
7. BMW Group. "BMW Group to deploy humanoid robots in production in Germany for the first time." press.bmwgroup.com, February 27, 2026. https://www.press.bmwgroup.com/global/article/detail/T0455864EN
8. The Humanoid Hub. "Figure's Helix model uses a System 2 (S2) and System 1 (S1) architecture." X (Twitter), February 2025. https://x.com/TheHumanoidHub/status/1892677115537195416
9. Humanoids Daily. "The End of C++: Brett Adcock on Helix 02 and Figure's Path to 'Room-Scale' Autonomy." humanoidsdaily.com, February 12, 2026. https://www.humanoidsdaily.com/news/the-end-of-c-brett-adcock-on-helix-02-and-figure-s-path-to-room-scale-autonomy
10. The Robot Report. "Figure humanoid robots use Helix VLA model to demonstrate household chores." therobotreport.com, February 22, 2025. https://www.therobotreport.com/figure-humanoid-robots-demonstrate-helix-model-household-chores/
11. Interesting Engineering. "Figure introduces 'never-seen' humanoid brain that controls full robot body." interestingengineering.com, February 20, 2025. https://interestingengineering.com/innovation/figure-launches-helix-ai-robots
12. New Atlas. "Figure's humanoids start doing tasks they weren't trained for." newatlas.com, February 21, 2025. https://newatlas.com/robotics/helix-vla-figure-02-robot/
13. NVIDIA. "Isaac GR00T N1: An Open Foundation Model for Generalist Humanoid Robots." research.nvidia.com. https://research.nvidia.com/publication/2025-03_nvidia-isaac-gr00t-n1-open-foundation-model-humanoid-robots
14. Physical Intelligence. "pi0: A Vision-Language-Action Flow Model for General Robot Control." arxiv.org. https://arxiv.org/html/2410.24164v1
15. Zitkovich et al. "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control." arxiv.org, 2023. https://arxiv.org/abs/2307.15818
16. TIME. "Figure 03: The Best Inventions of 2025." time.com. https://time.com/collections/best-inventions-2025/7318493/figure-03/
17. eWeek. "Figure Launches Humanoid Robot Upgrade Helix 02 for Full-Body Autonomy." eweek.com. https://www.eweek.com/news/figure-helix-02-humanoid-robot-autonomy/
18. Figure AI. "Scaling Helix: A New State of the Art in Humanoid Logistics." figure.ai, June 7, 2025. https://www.figure.ai/news/scaling-helix-logistics
19. Figure AI. "Helix Learns to Fold Laundry." figure.ai, August 12, 2025. https://www.figure.ai/news/helix-learns-to-fold-laundry
20. Figure AI. "Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot Transfer." figure.ai, September 18, 2025. https://www.figure.ai/news/project-go-big
21. Figure AI. "Figure Exceeds $1B in Series C Funding at $39B Post-Money Valuation." figure.ai, September 16, 2025. https://www.figure.ai/news/series-c
22. CNBC. "Figure AI sued by whistleblower who warned that startup's robots could 'fracture a human skull'." cnbc.com, November 21, 2025. https://www.cnbc.com/2025/11/21/figure-ai-sued.html
23. Humanoids Daily. "Figure Claims Production Milestone as 'BotQ' Ramps Up Figure 03 Manufacturing." humanoidsdaily.com, April 1, 2026. https://www.humanoidsdaily.com/news/figure-claims-production-milestone-as-botq-ramps-up-figure-03-manufacturing
24. Figure AI. "Ramping Figure 03 Production." figure.ai, April 29, 2026. https://www.figure.ai/news/ramping-figure-03-production
25. TechTimes. "Figure AI's Helix-02 Robots Complete Full 8-Hour Autonomous Shifts as Humanoid Race Intensifies." techtimes.com, May 14, 2026. https://www.techtimes.com/articles/316632/20260514/figure-ais-helix-02-robots-complete-full-8-hour-autonomous-shifts-humanoid-race-intensifies.htm
26. Interesting Engineering. "Figure AI humanoids sort 28,000 packages in 24-hour autonomous test." interestingengineering.com, May 14, 2026. https://interestingengineering.com/ai-robotics/figure-ai-humanoids-24-hour-autonomous-run
27. Humanoid Guide. "Figure 02 - Specs, Price and Capabilities." humanoid.guide. https://humanoid.guide/product/figure-02/
28. Figure AI. "Helix 2.5: Zero-Shot 30-Home Generalization." figure.ai, September 17, 2026. https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization
29. Figure AI. "Introducing Index: Building The World's Largest and Most Diverse Physical Dataset." figure.ai, August 25, 2026. https://www.figure.ai/news/introducing-index
30. Figure AI. "Helix 02 Living Room Tidy." figure.ai, March 9, 2026. https://www.figure.ai/news/helix-02-living-room-tidy
31. Figure AI. "Helix-02 Bedroom Tidy." figure.ai, May 8, 2026. https://www.figure.ai/news/helix-02-bedroom-tidy
32. Figure AI. "F.03 Arrives at BMW." figure.ai, June 30, 2026. https://www.figure.ai/news/f-03-at-bmw
33. Figure AI. "Figure and Nscale Sign Strategic Partnership For Up to 100,000 GPUs on the NVIDIA Vera Rubin Platform." figure.ai, September 3, 2026. https://www.figure.ai/news/figure-and-nscale-sign-strategic-partnership
34. Figure AI. "Helix Loads the Dishwasher." figure.ai, September 3, 2025. https://www.figure.ai/news/helix-loads-the-dishwasher
35. Humanoids Daily. "Figure's Helix 2.5 Takes on Chores in 30 Unseen Homes." humanoidsdaily.com, September 17, 2026. https://www.humanoidsdaily.com/news/figure-helix-2-5-30-unseen-homes
36. Humanoids Daily. "Figure AI Pops Champagne as Autonomous Marathon Crosses 200 Hours Without Hardware Failure." humanoidsdaily.com, May 22, 2026. https://www.humanoidsdaily.com/news/figure-ai-pops-champagne-as-autonomous-marathon-crosses-200-hours-without-hardware-failure
37. TechCrunch. "Figure drops OpenAI in favor of in-house models." techcrunch.com, February 4, 2025. https://techcrunch.com/2025/02/04/figure-drops-openai-in-favor-of-in-house-models/
38. Humanoids Daily. "Figure AI Reorganizes to Boost Humanoid Learning with 'Helix' AI Model." humanoidsdaily.com, May 2025. https://www.humanoidsdaily.com/news/figure-ai-reorganizes-to-boost-humanoid-learning-with-new-helix-ai-model
39. Figure AI. "Figure 03." figure.ai. https://www.figure.ai/figure
40. Figure AI. "Company." figure.ai. https://www.figure.ai/company
41. Humanoids Daily. "Beyond the 60-Hour Mark: Figure AI's Endurance Marathon Signals Playbook for Figure 4 and Supply Chain Independence." humanoidsdaily.com, May 16, 2026. https://www.humanoidsdaily.com/news/beyond-the-60-hour-mark-figure-ai-s-endurance-marathon-signals-playbook-for-figure-4-and-supply-chain-independence
42. Physical Intelligence. "openpi." github.com. https://github.com/Physical-Intelligence/openpi
43. NVIDIA. "NVIDIA Isaac GR00T N1.7: Open Reasoning VLA Model for Humanoid Robots." huggingface.co, April 17, 2026. https://huggingface.co/blog/nvidia/gr00t-n1-7
44. NVIDIA et al. "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots." arxiv.org, March 2025. https://arxiv.org/abs/2503.14734
45. Figure AI. "Figure Signs Agreement with Catalyst Brands to Scale Humanoid Operations." figure.ai, May 26, 2026. https://www.figure.ai/news/figure-signs-agreement-with-catalyst-brands
46. Figure AI. "F.02 Decommission." figure.ai, September 30, 2026. https://www.figure.ai/news/f-02-decommission