π0.5

π0.5 (also written pi0.5, pi 0.5, or π₀.₅, and pronounced "pi zero point five") is a vision-language-action model developed by the robotics company Physical Intelligence and released on April 22, 2025.^[1][2] It is a successor to the company's earlier π0 model, built to tackle one of the hardest problems in embodied AI: operating reliably in physical environments the robot has never seen during training, an objective Physical Intelligence calls open-world generalization.^[2] The headline demonstration showed a wheeled mobile manipulator cleaning kitchens and tidying bedrooms in entirely new homes that were not present in its training data, performing multi-stage tasks lasting 10 to 15 minutes.^[1][2]

π0.5 reaches this generalization through co-training on six heterogeneous data sources (robot demonstrations from many platforms, multimodal web data, object detections, verbal instructions, and high-level semantic subtask labels) and through a hierarchical, two-level inference scheme: the model first predicts a high-level natural-language subtask (for example, "pick up the mug"), then generates the low-level motor commands to carry it out.^[2] The paper's abstract states that the authors "demonstrate for the first time that an end-to-end learning-enabled robotic system can perform long-horizon and dexterous manipulation skills, such as cleaning a kitchen or bedroom, in entirely new homes."^[2]

The paper describing the model, "π0.5: a Vision-Language-Action Model with Open-World Generalization," was published simultaneously as an arXiv preprint (arXiv:2504.16054) and as a Physical Intelligence blog post on April 22, 2025.^[1][2] It subsequently appeared in the Proceedings of The 9th Conference on Robot Learning (CoRL 2025), PMLR Volume 305, pages 17 to 40.^[3] The paper lists 35 named authors from Physical Intelligence, including Kevin Black, Noah Brown, Chelsea Finn, Karol Hausman, Brian Ichter, Sergey Levine, Allen Z. Ren, and Homer Walke, among others.^[2]

What is π0.5?

π0.5 is a single neural network (a vision-language-action model of roughly 3.3 billion parameters) that maps a robot's camera images and a natural-language instruction directly to motor commands, with an explicit intermediate step that names the current subtask in plain language.^[2] It is made by Physical Intelligence, a San Francisco robot learning company, and it builds directly on the π0 architecture.^[1][2] Its central contribution is a co-training framework that mixes six heterogeneous data categories, including mobile manipulator household demonstrations, non-mobile multi-environment robot data, cross-embodiment laboratory recordings, high-level semantic subtask annotations, internet-scale web data, and verbal instruction demonstrations.^[2] In experiments, π0.5 approached the performance of models trained directly on test-environment data while evaluating in homes that were not part of its training data.^[2]

Physical Intelligence summarized the design rationale in its announcement: "By training our VLA model on a variety of different data sources, we can teach it not only how to physically perform diverse skills, but also how to understand the semantic context of each skill."^[1]

Background

Physical Intelligence

Physical Intelligence (abbreviated as π or pi) is a San Francisco robotics AI company founded in 2024.^[13] Its founders include Karol Hausman, Brian Ichter, Sergey Levine, and Chelsea Finn, researchers with backgrounds spanning Google Brain, Google DeepMind, UC Berkeley, and Stanford. The company focuses on developing foundation models for physical systems, pursuing the same scaling approach that produced general-purpose text and image models, applied to robot control. Physical Intelligence raised early funding from investors including Bezos Expeditions, OpenAI, Sequoia Capital, and Khosla Ventures.

The company's research strategy is to train a single large model on data from many different robot types and tasks, rather than building narrow task-specific policies. The hypothesis is that a model trained on diverse enough physical interaction data will develop representations that transfer across embodiments and environments, much as large language models trained on diverse text transfer across domains.

π0 and the VLA lineage

π0 was Physical Intelligence's first major model release, published in October 2024.^[4] The model combined a 3-billion-parameter PaliGemma vision-language backbone with a separately parameterized action expert.^[4] The action expert used flow matching, a technique from generative modeling that learns smooth vector fields to convert a Gaussian noise distribution into a target action distribution.^[4] During inference, the model runs 10 denoising steps to produce a 50-step action chunk covering one second of motor commands.^[4]

π0 demonstrated broad dexterity across tasks like laundry folding, box assembly, and table bussing.^[4] However, its generalization remained largely constrained to environments and object configurations seen during training. Evaluating it in genuinely new homes, with unfamiliar furniture arrangements and unseen objects, revealed substantial performance drops.

To extend generalization beyond the training distribution, Physical Intelligence pursued two parallel lines of follow-up work. One produced FAST (Frequency-based Action Sequence Tokenizer), which compresses robot trajectories into compact discrete token sequences using Discrete Cosine Transform, enabling more efficient training on heterogeneous data.^[8] The other produced π0.5, which incorporated FAST tokens into a broader co-training framework explicitly designed for open-world generalization.^[2]

The open-world generalization problem

Generalization in robotics has a specific meaning distinct from its use in machine learning benchmarks. A robot that "generalizes" in the standard ML sense may succeed on held-out test objects from the same distribution as training objects. Open-world generalization refers to a harder condition: operating in physical spaces with different floor plans, furniture arrangements, lighting conditions, and object inventories from anything seen during training.

Most VLA papers published before 2025 evaluated robots in environments that closely matched their training settings. Even when test objects differed, the room, the table positions, and the camera angles were familiar. π0.5 defined its evaluation around a stricter criterion: the test homes used for evaluation were not part of the training set.^[2] This framing shaped the paper's entire methodology.

When was π0.5 released and where was it published?

π0.5 was announced via a blog post on the Physical Intelligence website on April 22, 2025, accompanied by the arXiv preprint and demonstration videos showing the robot operating in homes that were not in its training data.^[1][2] The blog framed the work as "a significant step forward toward truly generalizable physical intelligence," emphasizing the departure from prior VLA evaluations in training-matched settings.^[1]

The release generated substantial discussion in the robotics and machine learning communities. The paper highlights that 97.6% of the training examples in the first training phase came from sources other than mobile manipulators performing household tasks.^[2] The model used to train a household mobile robot turned out to rely overwhelmingly on data from other robots, web images, and non-mobile manipulation setups, yet still generalized to mobile household tasks. The paper presents this as evidence that its co-training recipe enables highly generalizable control of a mobile manipulator with only a medium-sized mobile manipulation dataset.^[2]

The paper's acceptance to CoRL 2025, the premier conference for robot learning research, confirmed its standing in the field.^[3] The earlier π0 paper was published at Robotics: Science and Systems (RSS) 2025.^[15]

How is π0.5 built (architecture)?

PaliGemma backbone

π0.5 is built on PaliGemma, Google's open-weight vision-language model.^[2] PaliGemma combines two pretrained components: SigLIP-So400m, a contrastive image encoder trained with SigmoidLoss on large image-text datasets, and Gemma-2B, a 2-billion-parameter decoder-only language model.^[2] These are joined by a learned linear projection that maps SigLIP image patch tokens into the Gemma token embedding space.

PaliGemma processes images at 224x224 resolution, generating 256 image tokens per image. The model can accept multiple images in a single context, which is how π0.5 handles the robot's four-camera input: the paper uses all four cameras (forward, backward and both wrists) for high-level inference, and the wrist and forward cameras for low-level action inference.^[2]

The total parameter count for π0.5 is approximately 3.3 billion.^[2] PaliGemma accounts for roughly 3 billion of these parameters.^[2] A separate action expert module contributes approximately 300 million additional parameters initialized from scratch rather than from pretrained weights.^[2]

Action expert

The action expert is a transformer-based module that takes PaliGemma's contextualized token representations and generates continuous robot action sequences. It processes a combination of image embeddings, language instruction tokens, robot proprioceptive state tokens, and, during post-training, semantic subtask prediction tokens.^[2]

The action expert uses adaptive RMSNorm (AdaRMS) to inject the flow-matching diffusion timestep into each transformer layer.^[2] A two-layer MLP maps the timestep scalar to a set of scale and shift parameters that modulate the layer normalization applied to each hidden state. This is the primary architectural change from π0's action expert, which fused the timestep directly with the noisy action input rather than conditioning each layer independently.^[4]

At inference time, the action expert runs 10 denoising steps using the flow matching procedure.^[2] Starting from Gaussian noise, each step moves the action representation along a learned vector field toward the target action distribution. The result is a 50-step action chunk covering one second of motion at 50 Hz, representing target poses for the arms, grippers and torso lift and target velocities for the mobile base.^[2]

Dual action representation

π0.5 uses two distinct action representations in different training phases. During pretraining, robot demonstration data is encoded using FAST tokens.^[2][8] The FAST tokenizer applies a Discrete Cosine Transform to an action trajectory, compressing the high-frequency components and retaining the low-frequency ones, then quantizing the result into a discrete codebook.^[8] FAST tokens allow robot demonstration data to be treated as discrete sequences comparable to text tokens, enabling joint training with language and web data using a single cross-entropy objective.^[8]

During post-training, a randomly initialized action expert is added and trained with flow matching to produce continuous action chunks, while next-token prediction continues alongside it.^[2] The FAST token head is retained but the flow matching head becomes the primary output for robot control. This design means the model can leverage the computational efficiency of discrete token training at scale during pretraining, then recover continuous action precision during the post-training specialization phase.

How does π0.5 do hierarchical (high-level then low-level) inference?

One of the clearest architectural additions in π0.5 relative to π0 is the explicit high-level subtask prediction stage, which gives the model a two-level, hierarchical inference process.^[2] Before generating motor action tokens, the model first produces a short natural-language description of the current semantic subtask (for example, "pick up the mug" or "open the drawer").^[2] This text is generated by the Gemma language head using autoregressive decoding.

The predicted subtask text is then prepended to the action expert's input context before flow matching begins. This gives the action expert an explicit semantic anchor: rather than inferring the current goal entirely from visual observations, it receives a natural-language description of what it is about to do. The authors described this as a form of chain-of-thought reasoning applied to robot control.^[2] In the paper's framing, the model "first infers a high-level subtask" and only then predicts the low-level motor action, mirroring how a person might decide what to do before doing it.^[2]

Ablation experiments showed that including subtask annotation data during training provided benefits even when the predicted subtask was not passed to the action expert at inference time.^[2] The authors concluded that a significant portion of the benefit of explicit high-level inference is obtained simply by including subtask prediction data in the training mixture.^[2]

Attention structure

Like π0, π0.5 uses a prefix-suffix attention structure.^[2] Image, language, and state tokens form the prefix, over which bidirectional (full) attention is applied. Discrete FAST action tokens attend to the prefix and autoregressively to preceding action tokens, while the continuous action-expert tokens attend to the prefix and bidirectionally to one another. This structure ensures the model can integrate full context when generating actions while preserving the autoregressive structure needed for FAST token generation.

Training differences from π0

Several training-time choices distinguish π0.5 from π0. In π0, the PaliGemma backbone and the action expert were trained together, with the action expert receiving continuous flow-matching supervision throughout.^[4] In π0.5, the whole model is likewise trained, with web data included in both phases to preserve the model's semantic and visual capabilities.^[2]

Knowledge insulation, in which the gradient from the action expert is blocked from flowing back into the VLM backbone, was introduced in a separate May 2025 Physical Intelligence paper rather than in π0 or the original π0.5 paper.^[16] In the π0.5 paper, post-training jointly trains next-token prediction (to preserve text prediction) and the newly added flow-matching action expert.^[2] Some later openpi π0.5 checkpoints, such as π0.5-DROID, were trained with knowledge insulation.^[14]

How does co-training on heterogeneous data work?

Overview of co-training

The defining methodological contribution of π0.5 is training a single model jointly on six data categories that differ dramatically in their format, source robot, semantic content, and action space.^[2] The hypothesis driving this design is that different data sources teach complementary capabilities: mobile manipulation data teaches the target skill, non-mobile robot data teaches generalizable manipulation primitives, cross-embodiment data transfers motor knowledge across platforms, semantic annotation teaches the structure of household tasks, web data teaches object recognition and scene understanding, and verbal instruction data teaches language grounding. The paper's abstract states plainly that "this kind of knowledge transfer is essential for effective generalization."^[2]

Joint training on all six categories is not straightforward because the data types are heterogeneous: some examples include action sequences while others do not, some have bounding box annotations while others have only image-text pairs, and the robot embodiments have different observation and action spaces. The FAST tokenizer and the dual action representation solve part of this problem by providing a common discrete token format for action data.^[8] The remainder is handled through a shared token vocabulary and a multi-task loss that applies different output heads to different data types depending on what labels are available.^[2]

Data categories

The most striking aspect of the data composition is the relative scale of non-MM sources. During pretraining, 97.6% of training examples come from sources other than mobile manipulator household data.^[2] Cross-embodiment recordings, multi-environment robot demonstrations, semantic annotations, and web images collectively dominate the training distribution. Despite this, the final model generalizes to mobile household manipulation.^[2]

Mobile manipulator data (MM)

The MM data was collected by Physical Intelligence's data collection team operating the mobile manipulator hardware in approximately 100 distinct household environments.^[2] Demonstrations were collected via teleoperation, with human operators remotely controlling the robot's base, torso, arms, and grippers while a researcher indicated the task objective. The target tasks were the four evaluation categories: dishes in sink, items in drawer, laundry basket, and make bed.^[2] Approximately 400 hours of MM demonstrations were collected across the roughly 100 homes.^[2]

Even with 100 training homes, the MM data represents a small fraction of the total training volume. The paper's key empirical finding is that this relatively modest collection of on-robot household data, combined with diverse data from other sources, is sufficient to achieve near-parity with oracle baselines trained directly on the test environments.^[2]

Cross-embodiment laboratory data (CE)

The CE data includes two components. The first is a subset of the Open X-Embodiment (OXE) dataset, a publicly available corpus assembled by Google and academic collaborators that aggregates robot manipulation demonstrations from over 20 institutions and 22 robot platforms.^[2] The second is in-house cross-embodiment data collected by Physical Intelligence across multiple robot configurations, including single-arm manipulators and bimanual setups used for the original π0 training.^[2]

The OXE and in-house CE data provide manipulation demonstrations across a wide range of object types, task structures, and robot kinematics. Because these demonstrations are mostly from laboratory settings rather than household environments, they do not directly teach household task completion. Instead, they appear to teach generalizable manipulation skills: how to grasp objects of different shapes, how to handle soft materials, how to operate drawers and doors.

Web data (WD)

The web data consists of internet-scale image understanding examples: image captioning pairs, visual question-answering examples, and object localization examples with bounding box annotations.^[2] These examples do not include robot actions. They are included in training as a non-robotic co-training signal.

The motivation for including web data is to prevent catastrophic forgetting of the PaliGemma backbone's visual recognition capabilities. A model trained exclusively on robot demonstration data may lose the ability to recognize object categories not represented in robot demonstrations. Since the test homes contained arbitrary household objects not present in training demonstrations, the web data was expected to be critical for out-of-distribution object recognition.

Ablation experiments supported this: removing web data significantly worsened language following on out-of-distribution objects, while its effect on end-to-end task performance was not statistically significant.^[2]

High-level subtask annotation (HL)

The HL data consists of robot observation frames or short video clips paired with manually written natural-language descriptions of the current semantic subtask.^[2] These descriptions are typically short phrases ("pick up the plate," "walk to the kitchen") rather than full sentences. Annotation was performed by the data collection team after the fact, labeling segments of existing robot demonstrations with semantic subtask descriptions.

The HL data trains the model's high-level language head: the capacity to generate a short text description of what the robot is about to do before generating actions. Including HL data also provides a richer supervision signal for the visual representations, since the model must learn to associate specific visual configurations with specific semantic subtask labels.

Two-stage training procedure

Training proceeds in two sequential phases.

The pretraining phase runs for 280,000 gradient steps and uses five data categories (MM, ME, CE, HL and WD); verbal instruction (VI) data is added only in post-training.^[2] During this phase, all action data is represented using FAST tokens.^[2] The PaliGemma backbone is unfrozen so it can adapt to the robot-specific distribution while retaining general visual and language representations through the stabilizing influence of web and language data.^[2] The loss function is a weighted sum of cross-entropy over all discrete token outputs.

The post-training phase runs for an additional 80,000 gradient steps and focuses the model on mobile manipulation.^[2] This phase uses MM and ME robot data (filtered to successful episodes below a length threshold), web data (WD) to preserve semantic and visual capabilities, the multi-environment slice of HL data, and VI data.^[2] Training continues next-token prediction alongside the new action expert to preserve the model's text prediction capabilities.^[2] The flow matching action head is introduced during this phase, and its loss is combined with the cross-entropy loss at a weight of α = 10.0.^[2] Post-training specializes the action expert for the continuous action precision required by mobile manipulation tasks.

How does π0.5 generalize to new homes?

Definition and prior work context

The term "open-world generalization" in π0.5's framing refers to performance in physical environments not represented in the training data.^[2] Prior VLA evaluations, including those for π0 and most VLAs published before 2025, tested robots in environments matched to or very close to training conditions. Even evaluation on novel test objects typically occurred in familiar rooms with familiar furniture. The few prior works that did evaluate in new environments generally did so with simple pick-and-place tasks using a single robot arm, not with long-horizon mobile manipulation tasks.

For its most realistic final evaluation, π0.5 was tested in three real homes that were not part of the training set, while mock home environments were used for controlled, reproducible comparisons.^[2] The homes had different furniture, kitchen layouts, counter arrangements, clutter distributions, and object inventories from anything in the training set. The evaluation tasks involved multiple stages lasting about 2 to 5 minutes, while separate long-horizon demonstrations of cleaning an entire kitchen or bedroom ran 10 to 15 minutes.^[2]

Hardware platform

π0.5 was evaluated on two mobile manipulator platforms.^[2] Each robot has a wheeled holonomic base (allowing movement in any direction without turning), a torso lift, two 6-degree-of-freedom arms with parallel-jaw grippers, and four RGB cameras: one forward-facing mounted on the torso, one rear-facing, and one wrist-mounted camera on each arm.^[2] The combined state and action space has 18 to 19 degrees of freedom.^[2] Motor commands are tracked by a proportional-derivative controller at 50 Hz.^[2]

This hardware configuration is simpler than a humanoid robot but more capable than a fixed-base arm: the mobile base lets the robot navigate household spaces, the torso lift helps with varying counter and shelf heights, and the wrist cameras provide close-up views useful for grasping.

Evaluation tasks

Four primary tasks structured the evaluation:

Each task was evaluated across multiple trials per task per environment.^[2] The primary metrics were task progress (the fraction of relevant items or subtasks successfully completed in each trial) and a language-following rate, measured in a separate experiment, of how often the robot selected the object named in a specific command.^[2] A separate analysis tracked out-of-distribution generalization specifically to objects not present in any training demonstration.^[2]

Quantitative results

The central quantitative result is that, evaluated in homes the team had never collected data in, π0.5 approached the performance of an oracle baseline trained directly on data from the test environments.^[2] In other words, the co-training methodology produced nearly environment-matched performance without requiring any data collection in the target environments, which the paper presents as evidence that knowledge transfer from diverse sources is what drives generalization.^[2]

The paper also reported a scaling curve: as the number of distinct training homes in the MM dataset increased (up to roughly 100), task performance and language-following improved steadily, and generalization to new homes approached oracle levels at around 100 training environments.^[2] At a handful of training homes, generalization to new homes was poor.^[2] This curve indicates that environment diversity in the training data, not merely total demonstration volume, was the dominant factor in open-world performance.^[2]

Ablation studies

The paper reported ablations isolating each data source's contribution:^[2]

• Removing the CE (cross-embodiment) data significantly degraded performance, particularly on items in drawer and dishes in sink, suggesting that data from laboratory robot platforms transfers generalizable manipulation skills to the mobile household setting.^[2]
• Removing the ME (multi-environment non-mobile) data also hurt performance substantially, indicating that non-mobile demonstrations from diverse indoor environments contribute to spatial and object-level generalization beyond what the MM data alone provides.^[2]
• Removing the WD (web data) did not significantly change end-to-end task performance, but it significantly worsened language following on out-of-distribution objects and weakened high-level subtask inference.^[2]
• Removing HL data from training entirely (the 'no HL' ablation, which also has no high-level inference) performed significantly worse than the 'implicit HL' variant trained with HL data.^[2]
• An ablation that removed subtask prediction at inference time while retaining the subtask training data showed intermediate performance, confirming that both the training signal and the inference-time chain-of-thought are independently useful.^[2]

What did the π0.5 demonstrations show?

Physical Intelligence published demonstration videos alongside the paper showing π0.5 operating in previously unseen homes.^[1] The kitchen cleaning sequences showed the robot identifying scattered dishes on countertops and tables, grasping items one at a time, navigating to the sink, placing items in the sink, and then returning to find additional items.^[1] The bedroom sequences showed the robot picking up clothing from the floor, handling soft textiles, locating a laundry basket through visual search, and depositing items.^[1]

One demonstration sequence showed the robot's response to human interference. During a task, a person moved an object the robot had already handled. The robot perceived the changed state on its next visual scan and re-executed the relevant subtask, demonstrating closed-loop replanning rather than execution of a fixed predetermined sequence.^[1] The robot also handled spills using a sponge in one demonstration, a task requiring both tool use and surface contact estimation.^[1]

The blog post noted that the robot "often exhibits a hint of the flexibility and resourcefulness with which a person might approach a new challenge," adapting its approach based on the specific object arrangement it found in each room rather than following a memorized path through the space.^[1]

All demonstrations were conducted without prior teleoperation in the test homes.^[2]

How is π0.5 different from π0?

π0.5 extends π0 in several ways that reflect both architectural changes and methodological shifts in training:

In direct comparison experiments within the π0.5 paper, π0.5 substantially outperformed both the original π0 and a variant called π0-FAST+Flow (which added the FAST tokenizer and flow matching to π0 without the full co-training data mix and semantic subtask prediction).^[2] The advantage held even when π0 was trained for up to 300k steps; π0.5 differs from π0-FAST+Flow mainly in adding the HL and web data.^[2]

How does π0.5 compare to other VLA models?

π0.5 appeared alongside several other significant VLA model releases in early 2025. Two of the most prominent were Helix from Figure AI and GR00T N1 from NVIDIA.^[5][6]

The architecturally deepest distinction is how each system handles the decomposition between semantic reasoning and motor control. Helix and GR00T N1 both implement a hard two-system split, with separate model components for high-level reasoning (running at lower frequency) and low-level motor control (running at higher frequency).^[5][6][12] π0.5 implements a softer decomposition: a single unified transformer produces both semantic subtask text and motor action tokens, using different output heads for different modalities but sharing weights throughout.^[2][12] Whether the unified or split architecture is ultimately more capable remains an open research question.

Another distinction concerns embodiment. Helix is purpose-built for the Figure humanoid robot, with 35 degrees of freedom including individual finger control.^[5] GR00T N1 targets humanoid robots broadly.^[6] π0.5 was built and evaluated on a wheeled mobile manipulator without humanoid legs, making direct performance comparisons across platforms difficult since the platforms have different capabilities and face different challenges.

In terms of data strategy, GR00T N1 placed heavy emphasis on synthetic data (DexMimicGen simulation trajectories and video-model-generated neural trajectories) and egocentric human video from open datasets such as Ego4D (about 2,500 hours of human video in total).^[6] π0.5 did not use synthetic simulation data, relying instead on real robot demonstrations and web images.^[2] Helix trained on approximately 500 hours of human teleoperation data, much less than either competitor.^[5]

What are the limitations of π0.5?

The π0.5 paper acknowledged several persistent failure modes and scope limitations.^[2]

The model uses a relatively modest context and has no persistent memory; the authors suggest that richer context and memory would help with partial observability, such as tasks that require navigating between rooms or remembering where objects are stored.^[2] Tasks requiring the robot to track where it placed objects after moving away, or to remember the state of areas not in the current field of view, can cause failures.

The high-level subtask predictor can be distracted by salient but irrelevant objects.^[2] A visually prominent item that has already been handled may re-attract the model's attention and cause repeated approach attempts. This is a symptom of the model lacking an explicit world state representation: it must infer the current state from raw visual observations at each timestep rather than consulting a maintained task progress record.

The system handles simple and direct commands reliably but was not evaluated on complex conditional or ambiguous instructions.^[2] Performance on instructions like "put the dishes away unless the dishwasher is already full" or "clean the kitchen when you have time" was not characterized.

The paper also lists environment-specific failures, such as unfamiliar drawer handles and cabinets that are physically hard for the robot to open, and partial observability, such as the robot arm occluding a spill that should be wiped.^[2]

All evaluation was conducted in standard residential environments with ordinary household objects.^[2] Performance on non-standard appliances or living spaces that deviate from typical residential layouts was not characterized. Commercial kitchens, industrial settings, and outdoor environments were entirely outside the evaluation scope.

The evaluation was also limited in scale: a small number of test homes with a modest number of trials per task is a relatively small sample for establishing robust performance claims.^[2]

What came after π0.5?

Physical Intelligence followed π0.5 with π0.6 (written π₀.₆ or pi-star-0.6), released on November 17, 2025.^[7] The primary innovation in π*0.6 was a method called RECAP (RL with Experience and Corrections via Advantage-conditioned Policies), which incorporates online reinforcement learning so the robot can improve through experience gathered during real-world deployments rather than relying solely on offline demonstration data.^[7]

π0.6 was reported to more than double the throughput on some of the hardest tasks and to cut failure rates by a factor of two or more relative to imitation-only training, on tasks like laundry folding, espresso making, and box assembly.^[7] π0.6 is the RL-trained version of π0.6, the base model that succeeded π0.5.^[7]

Physical Intelligence also maintains the openpi GitHub repository (Apache 2.0), which first provided open weights for π0 and π0-FAST and, in September 2025, added π0.5 itself: a π0.5 base checkpoint for fine-tuning plus versions fine-tuned for the LIBERO benchmark and the DROID dataset.^[14] In April 2026 the company announced π0.7, which it describes as a steerable model with emergent capabilities.^[17]

ELI5: what is π0.5 in simple terms?

Imagine a robot with wheels and two arms that has only ever practiced cleaning in a set of training houses. π0.5 is the "brain" that lets it walk into a completely new house it has never seen, look around with its cameras, and still clean the kitchen or tidy the bedroom. It learned partly from videos of other robots and from ordinary pictures and captions on the internet, not just from cleaning houses. When it works, it first tells itself in plain words what to do next ("pick up the cup"), and then it moves its arms to do exactly that.^[1][2]

Continue reading

• PPhysical IntelligencePhysical Intelligence (also known as Pi or π) is an American artificial intelligence robotics company that builds general-purpose foundation models and learning algorithms designed to control any robot for any…Mentioned 21 times above
• PaliGemmaPaliGemma is an open vision-language model developed by Google that pairs the SigLIP image encoder with a Gemma language model, takes an image plus a text prompt as input, and produces text as output.Mentioned 14 times above
• FFlow MatchingFlow Matching is a simulation-free training framework for generative models that fits a time-dependent velocity field to transport samples from a source distribution (typically a standard Gaussian) to a data…Mentioned 10 times above
• Humanoid Robot"Humanoid" redirects here. For the London-based robotics company, see Humanoid (company).Mentioned 5 times above
• Helix (VLA model)Helix is a vision-language-action model (VLA) developed by Figure AI that controls humanoid robots by mapping camera images and natural language commands directly to continuous control.Mentioned 5 times above
• VVision-language-action modelA vision-language-action model (VLA model, or VLA) is a machine-learning model for robotics that uses visual observations and a natural-language instruction to generate actions a robot can execute.Mentioned 3 times above

More from See also

• π0
• Robot learning
• Embodied AI
• Isaac GR00T
• Open X-Embodiment

More to read

• Karol Hausman

Topics:AI ModelsEmbodied AIRobotics

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Physical Intelligence. "A VLA with Open-World Generalization." pi.website/blog/pi05. April 22, 2025.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53 ^54 ^55 ^56 ^57 ^58 ^59 ^60 ^61 ^62 ^63 ^64 ^65 ^66 ^67 ^68 ^69 ^70 ^71 ^72 ^73 ^74 ^75 ^76 ^77 ^78 ^79 ^80 ^81 ^82 ^83 ^84 ^85 ^86 ^87 ^88 ^89 ^90 ^91 ^92Physical Intelligence et al. "π0.5: a Vision-Language-Action Model with Open-World Generalization." arXiv:2504.16054. April 22, 2025.
3. ^1 ^2Physical Intelligence et al. "π0.5: a Vision-Language-Action Model with Open-World Generalization." Proceedings of The 9th Conference on Robot Learning, PMLR 305:17-40, 2025.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7Black, Kevin, et al. "π0: A Vision-Language-Action Flow Model for General Robot Control." arXiv:2410.24164. October 2024.
5. ^1 ^2 ^3 ^4Figure AI. "Helix: A Vision-Language-Action Model for Generalist Humanoid Control." figure.ai/news/helix. February 2025.
6. ^1 ^2 ^3 ^4NVIDIA. "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots." arXiv:2503.14734. March 2025.
7. ^1 ^2 ^3 ^4Physical Intelligence. "π*0.6: a VLA That Learns From Experience" and "π*0.6 Model Card." arXiv:2511.14759; website.pi-asset.com/pi06star/PI06_model_card.pdf. November 17, 2025.
8. ^1 ^2 ^3 ^4 ^5Physical Intelligence. "FAST: Efficient Robot Action Tokenization." pi.website/research/fast. 2025.
9. Kalil, Mike. "Physical Intelligence's π0.5 VLA with Open-World Generalization." mikekalil.com. 2025.
10. Edelman, Isaac. "A review of π0.5: Scaling Open-World Generalization in Vision-Language-Action Models." Toward Humanoids, Medium. 2026.
11. DeepWiki. "π₀ Model Family." deepwiki.com/Physical-Intelligence/openpi/4.2-p-model-family. 2025.
12. ^1 ^2Pebblous. "Three Teams, Three Robot Brains: A Head-to-Head Comparison of GR00T, Gemini, and π Architectures." blog.pebblous.ai. 2025.
13. ^The Robot Report. "Physical Intelligence raises $600M to advance robot foundation models." therobotreport.com/...ance-robot-foundation-models. November 2025.
14. ^1 ^2Physical Intelligence. "openpi" GitHub repository. github.com/...openpi. Accessed November 2026.
15. ^Black, Kevin, et al. "π0: A Vision-Language-Action Flow Model for General Robot Control." Proceedings of Robotics: Science and Systems XXI, 2025. roboticsproceedings.org/...p010
16. ^Driess, Danny, et al. "Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better." arXiv:2505.23705. May 2025.
17. ^Physical Intelligence. "π0.7: a Steerable Model with Emergent Capabilities." pi.website/...pi07. April 16, 2026.

Show all 17 references