# Dyna Robotics

> Source: https://aiwiki.ai/wiki/dyna_robotics
> Updated: 2026-09-30
> Fact-checked: 2026-09-30
> Categories: AI Companies, Embodied AI, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Dyna Robotics." aiwiki.ai, 30 Sept 2026. https://aiwiki.ai/wiki/dyna_robotics
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Dyna Robotics** (styled **DYNA Robotics** in its own materials) is an American embodied-AI and robotics startup, founded in 2024 and based in Redwood City, California, that builds [robot foundation model](https://aiwiki.ai/wiki/robot_foundation_model)s for commercial [robot manipulation](https://aiwiki.ai/wiki/robot_manipulation) and deploys them in real businesses such as laundromats, restaurants, hotels, and gyms.[1][2][6][34] It started with low-cost, stationary dual-arm stations that fold napkins and towels, and in September 2026 it added a wheeled semi-humanoid robot, Taku, as part of its Dyna-2.1 system.[1][20][21] Its first model, DYNA-1, was unveiled on April 29, 2025; Dyna marketed it as the first dexterous robot [foundation model](https://aiwiki.ai/wiki/foundation_models) deployed in commercial settings.[4] Its second-generation model, Dyna-2, announced in August 2026, is a [world action model](https://aiwiki.ai/wiki/world_action_model) pre-trained on more than one million hours of first-person human video.[19][26] Dyna was founded by Lindon Gao, York Yang, and Yecheng "Jason" Ma; Gao and Yang previously built the retail smart-cart company Caper AI, which Instacart acquired for $350 million in 2021.[2] By September 2025 the company had raised about $143.5 million across a seed round and a Series A, the latter reported to value it at more than $600 million.[6][7][8][11]

## What is Dyna Robotics?

Dyna Robotics is a foundation-model company for physical robots: it trains a single [embodied AI](https://aiwiki.ai/wiki/embodied_ai) policy that controls robot hardware to perform real manipulation work inside paying businesses. Its first commercial systems were pairs of inexpensive, fixed robotic arms; by September 2025 it had paired its model with custom hardware that adds a mobile base, and in 2026 it introduced a semi-humanoid.[1][10][20] The company marketed DYNA-1 as the first dexterous robot foundation model deployed in commercial settings, and it sells its robots as a subscription service rather than as standalone hardware.[4][11][32] Its stated long-term goal is "physical AGI," meaning general-purpose physical intelligence for robots.[6] Dyna describes itself as "a research-driven product company" and says it has hardware design, manufacturing, and production capability alongside its model research.[18][33]

## History

### When was Dyna Robotics founded?

Dyna Robotics was founded in 2024 (CEO Lindon Gao wrote that the company had just "turned one year old" when it closed its Series A in September 2025) and emerged from stealth on March 25, 2025, disclosing a $23.5 million seed round at the same time.[1][2][34] Fortune reported that Gao had left his role at Instacart and Caper about six months earlier; Gao told the magazine that the seed round valued Dyna at around $100 million and that the company already had 30 employees.[2] Its premise at launch was that the main barrier to robot adoption is cost rather than capability, so it focused on inexpensive pairs of stationary robot arms controlled by embodied AI models that each master one task or a narrow set of tasks at a time, beginning with chores like folding and food preparation.[1]

### Timeline

| Date | Event |
| --- | --- |
| March 25, 2025 | Emerges from stealth with a $23.5 million seed round co-led by CRV and First Round Capital[1][2] |
| April 29, 2025 | Announces DYNA-1 (Dynamism v1)[4] |
| September 15, 2025 | Announces a $120 million Series A[6] |
| Autumn 2025 | Live laundry-folding demos of the DYNA-1i model at the Actuate conference and at CoRL in South Korea[30] |
| October 2025 | Publishes a customer case study with Monster Laundry in Sacramento[32] |
| November 2025 | Reports a pre-trained DYNA-1 base model that folds laundry and sorts packages without task-specific post-training[31] |
| May 2026 | Co-founder York Yang publishes an essay on ten months of production deployments[33] |
| August 10, 2026 | Announces Dyna-2, a world action model pre-trained on more than one million hours of human video[26] |
| August 27, 2026 | Says Din Tai Fung is rolling out Dyna robots across its restaurant network[18] |
| September 29, 2026 | Launches Dyna-2.1 and the Taku semi-humanoid robot[20][21] |

### Who founded Dyna Robotics?

The three co-founders combine consumer-hardware operating experience with [robot learning](https://aiwiki.ai/wiki/robot_learning) research:

| Founder | Role | Background |
| --- | --- | --- |
| Lindon Gao | Co-founder and CEO | Co-founded Caper AI, an AI smart-cart company, sold to Instacart for $350 million (2021) |
| York Yang | Co-founder | Engineer and co-founder of Caper AI |
| Yecheng "Jason" Ma | Co-founder and Chief Scientist | PhD (2025) from the University of Pennsylvania GRASP Lab; robot-learning researcher; first author of the Eureka reward-design paper |

Gao and Yang are repeat founders: their previous company, Caper AI (also styled Caper Inc.), built computer-vision-enabled self-checkout shopping carts and was acquired by Instacart in 2021 for $350 million.[2][3] Ma completed his PhD at the University of Pennsylvania's GRASP Laboratory, advised by Dinesh Jayaraman and Osbert Bastani, and works on reinforcement learning and robot learning.[12] He was the first author of Eureka, a 2023 paper on using large language models to write reward functions for robot training.[2][35] Press releases describe him as a former [DeepMind](https://aiwiki.ai/wiki/google_deepmind) research scientist, and Salesforce Ventures describes him as a former researcher at DeepMind and NVIDIA.[1][6][10]

### How much funding has Dyna Robotics raised?

Dyna has raised roughly $143.5 million in two rounds.[6][11] The seed round was co-led by CRV and First Round Capital, and the Series A was led by RoboStrategy together with CRV and First Round Capital, with participation from several strategic corporate investors including the venture arms of NVIDIA, Amazon, [Samsung](https://aiwiki.ai/wiki/samsung), and LG.[1][5][6]

| Round | Date | Amount | Lead investors | Notable participants |
| --- | --- | --- | --- | --- |
| Seed | March 2025 | $23.5 million | CRV, First Round Capital | (co-led) |
| Series A | September 2025 | $120 million | RoboStrategy, CRV, First Round Capital | Salesforce Ventures, NVentures (NVIDIA), Amazon Industrial Innovation Fund, Samsung Next, LG Technology Ventures |

The Series A, announced on September 15, 2025, brought total funding to about $143.5 million and, according to Bloomberg, raised the company's valuation to more than $600 million; Investing.com and Sacra also reported the figure, with Sacra describing it as post-money.[6][7][8][11] Dyna said it would use the capital to expand its research and engineering teams and to accelerate development of its next-generation foundation model.[5][6][9][17] Announcing the round, co-founder and CEO Lindon Gao said, "A strong foundation model is key to scalable distribution. Our models continuously improve with each customer deployment, generating high-quality data. We are observing true generalization as our robot enters new environments; it simply works out of the box, with no additional data."[6] In a company blog post about the round, Gao wrote that Dyna ran "hundreds of model variants weekly across offline and online evals."[34] The strategic backers (NVIDIA, Amazon, Samsung, and LG corporate funds, plus enterprise-software investor [Salesforce](https://aiwiki.ai/wiki/salesforce) Ventures) gave Dyna ties to large hardware and cloud companies interested in [AI robotics](https://aiwiki.ai/wiki/ai_robotics).[5][6]

## What is DYNA-1?

DYNA-1 (short for Dynamism v1), announced on April 29, 2025, is Dyna's first robot foundation model.[3][4] Dyna's longer research write-up on the model is dated June 2025 (listed as June 24, 2025 on its research index).[29][36] The company positions it as a single-weight, general-purpose model intended to perform a variety of everyday manipulation tasks across different commercial environments rather than a model hand-tuned for a single workcell.[6] According to Sacra, it is delivered as a full-stack system: two industrial robotic arms on a compact wheeled base, with quick-swap grippers (suction, parallel, and custom end-effectors), designed to fit at an existing workstation and driven by the learned policy.[11] Dyna's later Dyna-2 report describes DYNA-1 as a vision-language-action ([VLA](https://aiwiki.ai/wiki/vision_language_action_model)) model with a mixture-of-transformers action component initialized from [Qwen3-VL](https://aiwiki.ai/wiki/qwen3_vl)-4B.[19] Describing the design philosophy, co-founder Jason Ma said, "Our first principle is to design robot foundation models that attain both generalization and performance."[6]

### What can DYNA-1 do? The headline demonstration

Dyna's launch centered on a continuous napkin-folding run. In its April 2025 press release, Dyna said a pair of stationary arms running DYNA-1 folded more than 800 napkins over 24 hours with no human intervention, at a 99.4 percent success rate and 60 percent of human throughput.[4] Its research write-up gives the count as "850+" napkins at about 60 percent of human speed, while early coverage in SiliconANGLE and Robotics & Automation News reported "more than 700."[3][16][29] The task is harder than it looks because it requires reliably separating a single napkin from a stack and recovering when several are pulled at once.[3][29] Dyna graded folds on a five-point scale and counted only grades 4 and 5 as production quality: 98 percent of folds reached grade 3 or better, but only 75 percent met that production bar.[29] Dyna said the underlying skills transferred to related tasks such as laundry folding and, at a client's request, cup-filling, and that the model folded napkins in a customer environment it had not been trained in, although quality and throughput dropped until it received some on-site training.[3][29]

### How is DYNA-1 trained?

DYNA-1 is trained with [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) and a custom [reward model](https://aiwiki.ai/wiki/reward_model).[3] The reward model estimates task progress on long dexterous tasks, which Dyna says lets the robot explore, recover from its own mistakes, and generate and curate its own training data; the company called it "the first scalable foundation reward model for robotics."[29] Dyna's write-up describes six weeks of "RM-in-the-loop" training in which the model went from failing after five minutes to sustained 24-hour runs of about 800 folds at production quality.[29] Because the model is deployed in production, customer deployments become additional training data: according to Sacra, robots stream sensor data to the reward model, which labels successful and failed manipulations to improve the network.[6][11] This data flywheel, where each deployment compounds the model's advantage, is central to the company's strategy and to its investors' thesis.[6][10]

### DYNA-1i and pre-training updates

In a research post listed as October 15, 2025 on its research index, Dyna described an improved model, DYNA-1i, trained on tens of hours of post-training data collected in its office. In 30-minute continuous shirt-folding trials it folded 22 shirts in seen environments and 20 in unseen ones (an office lobby, a parking lot, and the CoRL exhibition hall), about 40 shirts per hour in both cases. Dyna demonstrated it live at the Actuate conference and ran it for three days at the Conference on Robot Learning (CoRL) in South Korea on a new "Dynasaur" robot.[30] In November 2025, Dyna reported that its latest pre-trained DYNA-1 base model could fold laundry and sort packages in unseen environments without post-training, and could reach roughly 100 percent success on new tasks with as little as one hour of demonstration data.[31]

## What is Dyna-2?

Dyna-2 is Dyna's second-generation model, announced on August 10, 2026.[26][28] Dyna calls it a world action model (WAM): a single generative model, built on a video-diffusion backbone, that can denoise future video and future robot actions jointly or separately.[19] It was pre-trained on more than one million hours of egocentric human video, which Dyna equates to "roughly 170 years of continuous waking experience"; most of the footage is head-mounted recordings of people cooking, tidying, folding, and assembling, collected by data partners and Dyna's own operations.[19] Dyna's accompanying research post, "Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models," is dated August 15, 2026 on its research index.[19][36]

### Architecture

Architecturally, Dyna-2 is a mixture of transformers in which each modality (video and action) is tokenized separately and has its own set of [diffusion transformer](https://aiwiki.ai/wiki/diffusion_transformer) (DiT) layers, with proprioception fed directly to the action transformer. The action transformer is deliberately shallower and joins the video stream at early layers, which Dyna says cut real-time inference latency without hurting performance. The model is trained with [flow matching](https://aiwiki.ai/wiki/flow_matching). For the variant used in the scaling study, video and action losses share a trunk, so video prediction shapes the shared representation, but at inference time the policy neither generates nor attends to predicted future video.[19]

### Scaling-law results

Dyna trained the same architecture on nested subsets of exactly 1,000, 10,000, 100,000, and 1,000,000 hours of human data and reported the following, all from its own evaluations:[19]

| Question | Dyna's reported result |
| --- | --- |
| Held-out human data | Prediction error improved monotonically on all four metrics and fit a power law in hours; accuracy@0.1 rose 51 percent across the ladder, against 12 percent for MSE |
| Held-out robot data, zero-shot | On 39 tasks from two stationary bimanual YAM platforms (12 internal, 27 from xdof's ABC dataset) that the model never saw in pre-training, all metrics improved monotonically with human-data scale, with an inflection between 10,000 and 100,000 hours; Dyna calls this the first human-to-robot transfer scaling law |
| On-robot performance after post-training | Across 14 tasks, each with at most 10 hours of robot data, mean normalized score rose 20%, 28%, 45%, 53% across the four pre-training scales; the 1-million-hour model was best on 9 of 14 tasks |
| Threshold effects | No checkpoint up to 100,000 hours turned a lockbox key; the 1-million-hour model succeeded 90 percent of the time |
| Data efficiency | A bottle-cap untwisting task post-trained on about 10 minutes of robot data rose from 10 percent to 40 and 50 percent as pre-training scaled |
| Language following | A targeted drink-retrieval task, run on an early prototype of Dyna's semi-humanoid, rose from 58 percent to 83 percent |

The 14 on-robot tasks ran on three embodiments: 11 on a stationary bimanual platform with 6-DOF YAM arms and parallel-jaw grippers, two on the same arms fitted with WUJI-2 20-DOF dexterous hands, and one on the semi-humanoid prototype. Evaluators were not involved in model development.[19] Dyna also ran controlled comparisons and concluded that world modeling is what makes cross-embodiment transfer scale: at a fixed amount of action-labelled data, adding video-only human data improved robot-data predictions monotonically, which the company summarized as "video is the new scaling axis."[19] Dyna left compute and model-size scaling for future work.[19]

### Comparison with DYNA-1 and other capabilities

| Evaluation | Dyna's reported result |
| --- | --- |
| Early Dyna-2 vs DYNA-1 on 7 benchmark tasks, matched data and hyperparameters | 1.55 times DYNA-1's success rate and 1.12 times its quality grade; early Dyna-2 won 65 percent of head-to-head comparisons, DYNA-1 29 percent, with 6 percent tied |
| Zero-shot production pass rate at customer sites where both were deployed | 87 percent for Dyna-2 vs 46 percent for DYNA-1, graded by on-site operators against customer acceptance criteria, with the same post-training budget |
| Language-following benchmark | 35 percent with action-only pre-training, 67 percent with video co-training on the early corpus, 96 percent with video co-training on the full Dyna-2 corpus |
| One-step video generation | A distilled one-step student cut sampling for a three-second, three-view video from 10,203 ms to 110 ms on one H100 (about 90 times faster), at an FVD of 121 against 80 for the full teacher |

All of these figures come from Dyna's report rather than independent testing.[19] Dyna's press release describes some results differently from the research post: it cites 15 benchmark tasks where the post reports 14, 13 minutes of data for the bottle-cap task where the post says roughly 10 minutes, and a 133 percent improvement on instruction-following tasks; a corrected version of the release describes the 1.55 times figure as a customer-quality pass rate, while the research post reports it as a success rate.[19][26] Interesting Engineering's coverage repeated the press-release figures.[28]

### Training infrastructure

In a follow-up post, "Training Dyna-2 at million-hour scale, repeatably," Dyna described the data and training infrastructure behind the model. It said the million-hour dataset comprised 43 million episodes, stored in MCAP containers with H.264 video and topic-group chunking.[27]

| Area | Dyna's reported change |
| --- | --- |
| Storage | About 68 percent smaller than a per-frame JPEG baseline; sample reads about 2.9 times faster |
| Ingestion | From 14,000 to 440,000 episode-hours per week (31 times), after moving to Airflow DAGs with staggered starts and bin-packed batches |
| Training manifest | Time to first batch cut from about 48 hours to under a minute, using a data warehouse and memory-mapped columnar files |
| Data delivery | A cluster-local cache on node NVMe (built on Alluxio) serving about 2 GB/s per node, with 98 percent GPU utilization on a warm multi-node run |
| Optimizer | Topology-aware sharding of the [Muon](https://aiwiki.ai/wiki/muon_optimizer) optimizer, roughly 3 times faster at scale on [B200](https://aiwiki.ai/wiki/nvidia_b200) nodes |

## Dyna-2.1 and the Taku robot

On September 29, 2026, Dyna launched Dyna-2.1 (styled "DYNA 2.1" in its press release), which it calls a "physical agent" for whole workflows rather than single tasks.[20][21] The system centers on Taku, a new semi-humanoid robot whose name comes from the Japanese word *takumi* (master craftsman). Taku has a human-shaped upper body, a folding lower body for reaching low and high, a base on four steerable wheels, and two 7-degree-of-freedom arms.[20] Three model layers control it: a whole-body controller trained with reinforcement learning in simulation that runs at 100 Hz, an improved version of the Dyna-2 world action model that produces whole-body target trajectories, and a vision-language workflow orchestrator that tracks the job, keeps a text memory, and decides the next step.[20] Dyna said the controller was trained with thousands of simulated Taku robots running in parallel in [NVIDIA Isaac Sim](https://aiwiki.ai/wiki/nvidia_isaac_sim), and NVIDIA's robotics account promoted the launch on that basis.[20][25]

The launch demonstration was a one-hour hotel laundry-room workflow that Dyna described as uncut: loading and starting washers and dryers, unloading them, folding towels, and shelving the stacks.[20][23] Dyna called Dyna-2.1 "the first Physical Agent that achieves reliable super long-horizon whole-body autonomy," a company claim.[23] Launch materials describe its deployment status differently. The press release said "DYNA 2.1 is being deployed in hotels, laundromats, and restaurants," and ABC7 San Francisco reported that the company said the technology is "already being used" in hotels, restaurants and laundromats; the research post and Dyna's launch thread on X said the "next milestone" is to take the new system to real-world customer sites.[20][21][22][24] The press release did not name a customer site using Taku.[21] Details of the system are covered in the [Dyna-2.1](https://aiwiki.ai/wiki/dyna_2_1) article.

## How does Dyna make money? Business model

Dyna operates a business-to-business Robots-as-a-Service (RaaS) model. Rather than selling hardware outright, it charges a monthly fee per robot that, according to Sacra, bundles the hardware, software, maintenance, and continuous model updates into a single payment; Dyna's Monster Laundry case study also describes a RaaS subscription.[11][32] The approach turns a large up-front capital expense into an operating expense, which lowers the barrier for smaller businesses, and it gives Dyna recurring revenue plus a continuous stream of training data.[11] The founders have stressed that the hardware is deliberately cheap: Fortune reported in March 2025 that where many AI-powered robots cost hundreds of thousands of dollars, Dyna's were expected to cost tens of thousands of dollars when sold.[2]

## Where are Dyna's robots deployed?

Dyna says its robots moved into paying customer sites within months of launch. When it announced its Series A in September 2025, the company said its robots were running sixteen hours a day at hotels, restaurants, laundromats, and gyms.[5][6]

The most visible early deployment is at Monster Laundry in Sacramento, which describes itself as the first laundry center in North America to debut a robotic folding system from Dyna Robotics, nicknamed "Sophy Swiftfold."[14] The robot folds towels, linens, and clothing using arms, sensors, and vision; the operator framed it as a labor aid rather than a replacement, and ABC10 in Sacramento covered the installation.[14][15] In a case study Dyna published in October 2025, Dyna and Monster Laundry said the robot had folded over 200,000 towels from 10 commercial customers in its first three months at a 99 percent quality acceptance rate, and that Monster had increased commercial order capacity by 25 percent.[32]

In May 2026, co-founder York Yang wrote that Dyna's longest-running deployment had reached about ten months of daily use, but that getting each customer from "sale" to "running reliably without us" had taken "weeks-to-months of on-site engineering," most of which did not transfer to the next customer.[33] Dyna later said that in its stationary folding deployments a new site could reach its production bar in as little as three days, down from weeks to months of on-site engineering.[18][20]

## Din Tai Fung rollout announcement (August 2026)

On August 27, 2026, Dyna announced that Din Tai Fung, which it called "the highest revenue per location restaurant chain in the US," was rolling out Dyna robots across its restaurant network after the deployment crossed what Dyna called its return-on-investment threshold. The announcement did not give the number of robots or restaurants, and the ROI assessment was Dyna's own.[18] Dyna described it, "to the best of our knowledge," as the first time a robot foundation model company had taken a manipulation commercial pilot to scaled deployment.[18]

The task is napkin folding. Dyna said each robot needs to produce 1,500 table-ready napkins in an 18-hour shift to keep a Din Tai Fung dining room supplied. Dyna said that when it first launched in 2025, DYNA-1 was producing about 480 such napkins, folding roughly 35 an hour with 75 percent clearing the quality bar. Dyna said Dyna-2 folds 95 napkins an hour with 93 percent meeting the bar, "which comes to 1,590 table-ready napkins a day."[18] Dyna also said Dyna-2's language following let it place each folded napkin into one of ten designated stacks in a bin based on an instruction, rather than hard-coding bin positions for each site.[18] These are Dyna's production figures graded against the customer's criteria; the post did not publish an independent audit. In its separate Dyna-2 report, the company reported the 87 percent versus 46 percent zero-shot production pass rate for Dyna-2 and DYNA-1 across unnamed customer sites.[19]

The company said its deployed fleet generated more than one terabyte of raw data per day. Its robot runtime keeps a fixed-size, in-memory window of recent activity and promotes the window around an alert or operator-flagged condition into a replayable episode, and an automatic labeling system breaks episodes into steps and failure modes. Dyna gave the example of a Din Tai Fung site whose throughput slipped because of missed napkin grabs in the first step of the procedure.[18] Dyna also said its deployment process could reach a customer's production ROI bar in as little as three days; that is the company's shortest reported case, not a guaranteed deployment time.[18]

Dyna said the Din Tai Fung expansion, together with rollouts at hotels, logistics companies, and data centers, would bring its deployed fleet to hundreds of robots by the first half of 2027. This was a future fleet target, not a count of robots operating at the time of the announcement.[18]

## How well do Dyna's robots actually work? Reception

In a February 2026 review of real-world robot autonomy, the research group [Epoch AI](https://aiwiki.ai/wiki/epoch_ai) wrote in its laundry-folding assessment that DYNA Robotics "has the strongest evidence of sustained real-world operation," citing the 24-hour run of 850-plus napkins at about 60 percent of human speed with a 99.4 percent success rate and zero interventions.[13] Epoch also cited the Monster Laundry deployment, where the robot folded over 200,000 towels across ten commercial clients in three months with a 99 percent quality-acceptance rate, and DYNA-1i's roughly 40 shirts per hour.[13] Epoch added a caveat: Dyna's commercial deployments are limited to simple, uniform items such as towels and napkins, not the diverse clothing a household robot would have to handle, so the results show narrow, reliable specialization rather than broad household generality.[13] Fortune placed Dyna's approach in contrast to companies building general-purpose models, such as [Physical Intelligence](https://aiwiki.ai/wiki/physical_intelligence) and [Skild AI](https://aiwiki.ai/wiki/skild_ai), and humanoid hardware makers such as [Figure AI](https://aiwiki.ai/wiki/figure_ai) and [Agility Robotics](https://aiwiki.ai/wiki/agility_robotics); Sacra lists [Covariant](https://aiwiki.ai/wiki/covariant) and its [RFM-1](https://aiwiki.ai/wiki/rfm_1) model among Dyna's competitors.[2][11]

Dyna's own writing has been skeptical of parts of the robotics market: York Yang's May 2026 essay argued that large parts of the market were in a bubble because "the timeline is longer than current funding levels assume," and that deployment speed, not demos, was the test that mattered.[33]

## Related

- [embodied_ai](https://aiwiki.ai/wiki/embodied_ai)
- [robot_foundation_model](https://aiwiki.ai/wiki/robot_foundation_model)
- [world_action_model](https://aiwiki.ai/wiki/world_action_model)
- [dyna_2_1](https://aiwiki.ai/wiki/dyna_2_1)
- [robot_manipulation](https://aiwiki.ai/wiki/robot_manipulation)
- [physical_intelligence](https://aiwiki.ai/wiki/physical_intelligence)
- [covariant](https://aiwiki.ai/wiki/covariant)
- [robot_learning](https://aiwiki.ai/wiki/robot_learning)

## References

1. "Dyna Robotics Raises $23.5 Million to Commercialize Embodied AI with Low-Cost Robots." PR Newswire, March 25, 2025. https://www.prnewswire.com/news-releases/dyna-robotics-raises-23-5-million-to-commercialize-embodied-ai-with-low-cost-robots-302410263.html
2. "After Instacart exit, Caper AI founder launches robotics startup to automate dirty, dull, and dangerous work." Fortune, March 25, 2025. https://fortune.com/2025/03/25/exclusive-instacart-smart-cart-startup-350m-google-deepmind-low-cost-robots/
3. "Dyna Robotics debuts DYNA-1 foundation model for powering robots." SiliconANGLE, April 29, 2025. https://siliconangle.com/2025/04/29/dyna-robotics-debuts-dyna-1-foundation-model-powering-robots/
4. "Dyna Robotics Unveils DYNA-1: The First Commercial-Ready Robot Foundation Model Offering Fully Autonomous, Round-the-Clock Dexterity." PR Newswire, April 29, 2025. https://www.prnewswire.com/news-releases/dyna-robotics-unveils-dyna-1-the-first-commercial-ready-robot-foundation-model-offering-fully-autonomous-round-the-clock-dexterity-302441437.html
5. "Dyna Robotics closes $120M funding round to scale robotics foundation model." The Robot Report, September 16, 2025. https://www.therobotreport.com/dyna-robotics-closes-120m-funding-round-to-scale-robotics-foundation-model/
6. "Dyna Robotics Raises $120 Million to Advance Robotic Foundation Models on the Path to Physical Artificial General Intelligence." PR Newswire, September 15, 2025. https://www.prnewswire.com/news-releases/dyna-robotics-raises-120-million-to-advance-robotic-foundation-models-on-the-path-to-physical-artificial-general-intelligence-302556817.html
7. "Dyna Robotics Raises $120 Million in Funding From Nvidia, Amazon." Bloomberg, September 15, 2025. https://www.bloomberg.com/news/articles/2025-09-15/dyna-robotics-raises-120-million-in-funding-from-nvidia-amazon
8. "Dyna Robotics raises $120 million, valuation tops $600 million." Investing.com, September 2025. https://www.investing.com/news/company-news/dyna-robotics-raises-120-million-valuation-tops-600-million-93CH-4239224
9. "Dyna Robotics Raises $120M Series A to Advance Physical AGI." Built In San Francisco, September 18, 2025. https://www.builtinsf.com/articles/dyna-robotics-raises-120-million-series-a-20250918
10. "Welcome, Dyna Robotics!" Salesforce Ventures, 2025. https://salesforceventures.com/perspectives/welcome-dyna-robotics/
11. "Dyna Robotics valuation, funding & news." Sacra, 2025. https://sacra.com/c/dyna-robotics/
12. "Yecheng (Jason) Ma." University of Pennsylvania GRASP Lab. https://www.grasp.upenn.edu/people/yecheng-jason-ma/
13. Rivière, Yann, and Jean-Stanislas Denain. "Where Autonomy Works: Evaluating Robot Capabilities in 2026." Epoch AI, February 10, 2026. https://epoch.ai/blog/where-autonomy-works-evaluating-robot-capabilities-in-2026
14. "Robotic Folding Laundromat in Sacramento." Monster Laundry. https://www.monsterlaundry.com/robotic-folding-laundromat/
15. "Sacramento laundromat uses robot to fold clothes at Monster Laundry." ABC10 / YouTube. https://www.youtube.com/watch?v=w9OTGFX0hBY
16. "Dyna Robotics Unveils Dyna-1: Autonomous robot achieves 24-hour dexterous task performance." Robotics & Automation News, May 1, 2025. https://roboticsandautomationnews.com/2025/05/01/dyna-robotics-unveils-breakthrough-in-robust-real-world-embodied-ai/90152/
17. "Dyna Robotics banks $120M from Nvidia, Amazon to bring embodied AI to the real world." Tech Funding News, September 17, 2025. https://techfundingnews.com/dyna-robotics-banks-120m-from-nvidia-amazon-to-bring-embodied-ai-to-the-real-world/
18. Dyna Robotics. "Not Just a Model, But a Product." August 27, 2026. https://www.dyna.co/news/scaling-customer-deployments
19. Dyna Robotics. "Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models." August 2026. https://www.dyna.co/dyna-2
20. Dyna Robotics. "Dyna-2.1: A Physical Agent for End-to-End Workflows." September 29, 2026. https://www.dyna.co/dyna-2.1
21. DYNA Robotics. "DYNA Robotics Launches DYNA 2.1 Physical Agent, a Semi-humanoid Robot that Completes Full Workflows such as a Commercial Laundry Shift." PR Newswire, September 29, 2026. https://www.prnewswire.com/news-releases/dyna-robotics-launches-dyna-2-1-physical-agent-a-semi-humanoid-robot-that-completes-full-workflows-such-as-a-commercial-laundry-shift-302892411.html
22. "Meet Taku: Bay Area company's AI robot loads washers, folds towels and stocks shelves." ABC7 San Francisco (KGO-TV), September 29, 2026. https://abc7news.com/post/meet-taku-redwood-citys-dyna-robotics-ai-robot-loads-washers-folds-towels-stocks-shelves/19888144/
23. Dyna Robotics (@DynaRobotics). "We are releasing Dyna-2.1, the first Physical Agent that achieves reliable super long-horizon whole-body autonomy..." X, September 29, 2026. https://x.com/DynaRobotics/status/2104970523726033387
24. Dyna Robotics (@DynaRobotics). "Dyna-2.1 makes significant strides in making general-purpose robots useful in the real world..." X, September 29, 2026. https://x.com/DynaRobotics/status/2104971005865427434
25. NVIDIA Robotics (@NVIDIARobotics). "From simulation to real-world autonomy. Dyna Robotics trained Dyna-2.1 in part with NVIDIA Isaac Sim..." X, September 29, 2026. https://x.com/NVIDIARobotics/status/2105081108719440013
26. DYNA Robotics. "Dyna Robotics unveils DYNA-2 World-Action Model, demonstrating first true scaling law in robotics powered entirely by human data" (corrected release). PR Newswire, August 10, 2026. https://www.prnewswire.com/news-releases/dyna-robotics-unveils-dyna-2-world-action-model-demonstrating-first-true-scaling-law-in-robotics-powered-entirely-by-human-data-302847114.html
27. Dyna Robotics. "Training Dyna-2 at million-hour scale, repeatably." August 2026. https://www.dyna.co/research/dyna-2-infrastructure
28. Walter, Neetika. "Humanoid robots trained on 1M hours of human video achieve up to 90% task success." Interesting Engineering, August 10, 2026. https://interestingengineering.com/ai-robotics/dyna-robotics-dyna-2-human-video-robot-training
29. Dyna Robotics. "Dynamism v1 (DYNA-1) Model: A Breakthrough in Performance and Production-Ready Embodied AI." June 2025. https://www.dyna.co/research/dyna-1
30. Dyna Robotics. "Open-World Dexterity and Live Demos around the World." 2025. https://www.dyna.co/research/open-world-dexterity
31. Dyna Robotics. "DYNA-1 Pre-Training: Zero-Shot Dexterity Is Here." November 2025. https://www.dyna.co/research/pre-training
32. Dyna Robotics. "Customer Testimonial: Monster Laundry" (How Monster Laundry Scaled Its Laundromat Operation with DYNA Robotics). October 2025. https://www.dyna.co/news/monster-laundry
33. Yang, York. "What 10 Months in Production Taught Us About the Robotics 'Bubble'." Dyna Robotics, May 2026. https://www.dyna.co/news/robotics-bubble
34. Gao, Lindon. "DYNA Robotics Closes $120M Series A: How We Think About Scaling Robotic Foundation Models." Dyna Robotics, September 2025. https://www.dyna.co/news/series-a
35. Ma, Yecheng Jason, et al. "Eureka: Human-Level Reward Design via Coding Large Language Models." arXiv:2310.12931, 2023. https://arxiv.org/abs/2310.12931
36. Dyna Robotics. "Research" (index of research posts). https://www.dyna.co/research