The chips and AI inside the robots

2026 early-access GR00T release, 3 billion parameters, built on the Cosmos-Reason2-2B vision-language backbone. Unlike N1.6 its weights are offered under the NVIDIA Open Model License for commercial use. NVIDIA publishes optimised inference timings for Jetson Orin, Jetson AGX Thor and DGX Spark. It is the model family behind NVIDIA's Isaac GR00T reference humanoid announced in May/June 2026.

Used in: Unitree H2 Plus (NVIDIA Isaac GR00T Reference Humanoid), G1

Open reasoning driving model announced at CES in January 2026: a 10-billion-parameter vision-language-action model that takes camera video, outputs a driving path and also explains its reasoning. NVIDIA released the weights on Hugging Face with an open simulator and a driving dataset, and pitched it partly as a teacher model that carmakers adapt. The Mercedes-Benz CLA was announced as the first car to ship with NVIDIA's driving stack built around it, as an enhanced Level 2 system.

Used in: Mercedes-Benz CLA (2026, enhanced Level 2)

April 2026 'steerable' generalist model. It accepts several kinds of prompts (language, speed and quality hints, control-mode labels and picture subgoals) and is reported to combine skills it already knows to do tasks it was never trained on, such as folding shirts on a robot arm that had no shirt-folding data.

Used in: Universal Robots UR5e (shirt-folding example reported in coverage)

January 2026 upgrade that lets one network control the whole robot, not just the upper body. It adds 'System 0', a learned whole-body controller trained on more than 1,000 hours of human motion and running at 1 kHz for balance and contact. It uses Figure 03's palm cameras and touch sensing. In September 2026 Figure described Helix 2.5, pretrained on its Index human-behaviour dataset, tidying, folding and bed-making in homes it had not seen.

Used in: Figure 03

Third generation, announced 30 July 2026. Reporting says the action model now drives a whole humanoid, walking and crouching as well as using its hands, where earlier versions mainly handled the upper body. ER 2 adds live video understanding and multi-robot coordination and is available to developers as a preview through the Gemini API; the action and on-device models are for trusted testers. DeepMind lists Agile Robots, Apptronik and Boston Dynamics as partners; Boston Dynamics agreed in January 2026 to put Gemini Robotics models into the new Atlas.

Used in: Atlas (product version)

NVIDIA's first open foundation model for humanoids, released at GTC in March 2025. It splits the job in two: a slower vision-language part reads the scene and the instruction and plans, and a fast action part turns that plan into smooth joint movements. Robot makers download it and fine-tune it on their own real or simulated data. The published checkpoint is about 2 billion parameters (GR00T-N1-2B).

Used in: NEO Gamma

June 2025 update to GR00T N1. The vision-language backbone is kept frozen, grounding was improved and a new training objective lets the model learn from human videos as well as robot data. NVIDIA reports better instruction following and success rates in simulation and on real robots. About 3 billion parameters in total.

Used in: G1

Later 2025 GR00T release (3 billion parameters). It changes how the vision-language features connect to the action generator, trains jointly on action and world-modelling objectives and copes with a different number of cameras per robot. Trained on bimanual, semi-humanoid and humanoid data, real and synthetic. Weights were published under an NVIDIA licence that the model card links as non-commercial.

A family of generative world models launched at CES in January 2025. Developers use them to generate physically plausible video for training and testing robots and self-driving cars, or fine-tune them for their own use. Released under NVIDIA's open model licence. These models are used to make training data; they do not drive the robot directly.

Used in: 1X Technologies (maker-level adopter named by NVIDIA), Agility Robotics (maker-level adopter named by NVIDIA), Figure AI (maker-level adopter named by NVIDIA), XPENG (maker-level adopter named by NVIDIA), Uber and Waabi (autonomous-driving adopters named by NVIDIA)

April 2025 successor to π0 aimed at working in places the robot has never seen. It is trained on a mix of data from several robot types, high-level task labels and web data, and was shown cleaning kitchens and bedrooms in homes that were not in its training data. Its checkpoints were added to the open openpi repository in September 2025.

Used in: Franka arm in the DROID setup (π0.5-DROID checkpoint)

November 2025 model that keeps improving after deployment. Using a method the company calls RECAP, the robot learns from its own attempts, from corrections by human operators and from a learned estimate of how well each attempt went, which is a form of reinforcement learning on real hardware.

Figure's in-house vision-language-action model, announced in February 2025 shortly after Figure ended its OpenAI collaboration. A 7-billion-parameter vision-language model handles understanding at 7-9 Hz and an 80-million-parameter policy turns that into upper-body motion at 200 Hz. Each part runs on its own low-power embedded GPU inside the robot. Figure later showed it on logistics parcel sorting, laundry folding and dishwasher loading.

Used in: Figure 02

March 2025 pair of models built on Gemini 2.0. Gemini Robotics outputs robot actions directly; Gemini Robotics-ER is a spatial-reasoning model that robot makers can connect to their own controllers. Trained mainly on the two-arm ALOHA 2 platform and also run on a Franka two-arm setup and Apptronik's Apollo humanoid. Early testers included Agile Robots, Agility Robotics, Boston Dynamics and Enchanted Tools.

Used in: Apollo, ALOHA 2 bimanual robot, Franka bimanual platform

June 2025 version of Gemini Robotics small enough to run entirely on the robot, without a network connection. Trained on ALOHA, then adapted to a Franka FR3 two-arm robot and the Apptronik Apollo humanoid. Offered to trusted testers with an SDK; weights are not public.

Used in: Apollo, Franka FR3 bimanual robot

September 2025 models. Gemini Robotics 1.5 can think through a step in plain language before acting and transfer motions learned on one robot body to another. ER 1.5 acts as a planner that breaks jobs into steps and can call tools such as web search. ER 1.5 was opened to developers through the Gemini API; the action model stayed partner-only.

Used in: Apollo, ALOHA 2, Franka bimanual robot

June 2025 vision-language transformer for household mobile manipulation: fetching things, opening doors and moving around a home. At 160 million parameters it is small enough to run fully on NEO's embedded GPU at about 5 Hz. 1X says it also drives autonomy on its older EVE robot.

Used in: NEO Gamma, EVE

Skild calls this an 'omni-bodied' brain: one model meant to control quadrupeds, humanoids, table-top arms and mobile manipulators. A slow high-level policy plans manipulation and navigation and a fast low-level policy turns that into joint commands; it is pretrained on simulation and human video. In March 2026 ABB Robotics, Universal Robots and MiR agreed to integrate it, with a planned Foxconn line assembling NVIDIA servers. In August 2026 Skild introduced S1, which copies a task from a single video demonstration.

Used in: ABB Robotics industrial arms (integration partner, robots not named), Universal Robots cobots (integration partner)

August 2025 joint result: a 450-million-parameter diffusion transformer that takes language instructions and controls the electric Atlas end to end, combining walking with long chains of manipulation. It runs at 30 Hz and plans about 1.6 seconds of motion at a time. Training data came partly from an upper-body Atlas test stand.

Used in: Atlas (electric prototype)

AgiBot's March 2025 generalist model. AgiBot calls the design ViLLA: a vision-language model plus a 'latent planner' that learns from many robot types and human videos, and an action expert trained on more than a million real demonstrations. It was pretrained on data from AgiBot's G1 robot and reported to transfer to other platforms. AgiBot released it on GitHub in September 2025, after opening its AgiBot World dataset.

Used in: AGIBOT Genie G1

Open world-model-plus-action framework that Unitree published in September 2025. The world model learns how a robot's actions change its surroundings; it can act as a simulator that generates training data, or feed an action head that decides what to do. Code and weights are on GitHub with demos on Unitree hardware.

Used in: G1, Unitree Z1 robot arm

June 2025 world model of 1.2 billion parameters, trained mostly on video, that learns to predict what happens next in a scene. After extra training on robot data from the open DROID dataset it was used to plan robot-arm actions in Meta's labs. Meta released code and checkpoints for research and commercial use.

Used in: Robot arms in Meta's labs (trained on DROID data)

February 2025 model for AI agents that act in both software and the physical world: it understands images and video and produces plans and actions for user interfaces and robot arms. The 8-billion-parameter checkpoint (Magma-8B) is MIT-licensed. Robot results were reported on simulation benchmarks rather than a named commercial robot.

Compact open vision-language-action model of about 450 million parameters, released in June 2025 with full training recipes. It was trained only on community datasets and is meant to run on consumer hardware and cheap arms.

Used in: SO-100 arm, SO-101 arm, LeKiwi mobile robot

Described by Waymo in December 2025 as the world model behind the Waymo Driver. A fast encoder fuses camera, lidar and radar for quick reactions, and a Gemini-trained driving vision-language model reasons about rare situations. Both feed a decoder that predicts other road users, builds maps and generates and checks trajectories.

Used in: Waymo Jaguar I-PACE (5th-gen Waymo Driver), Waymo Ojai (6th-gen Waymo Driver, Zeekr-built)

Shown at XPeng's AI Day in November 2025 as one model family for its cars' driver assistance, its planned robotaxis, its IRON humanoid and its flying cars. XPeng says it goes straight from vision to action through hidden tokens, skipping a language step. It runs on XPeng's own Turing chips; IRON carries three of them. XPeng said partners such as Volkswagen would get access, but no public weights were confirmed.

Used in: IRON (next generation), XPeng robotaxi (announced November 2025)

A general robot-control model that pairs a vision-language model with a 'flow matching' action generator so it can output smooth, fast motions. Shown in October 2024 on single-arm, two-arm and mobile robots doing tasks like folding laundry. In early 2025 Physical Intelligence released the weights and code as openpi under Apache 2.0, together with π0-FAST, a variant that turns actions into tokens.

Used in: ALOHA bimanual robots (fine-tuned checkpoints in openpi), Franka arm in the DROID setup, Universal Robots UR5 (example in openpi)

In February 2024 Figure announced a deal with OpenAI to build AI models for its humanoids; OpenAI's models gave Figure 01 and Figure 02 spoken conversation and language reasoning. Figure ended the agreement in February 2025 to build its own models (Helix). OpenAI earlier did robot-hand research (Dactyl, about 2018-2019) and in 2026 was reported to be rebuilding a robotics hardware team; no OpenAI robot model has been released publicly.

Used in: Figure 01, Figure 02

September 2024 learned simulator: given what the robot sees and what it is about to do, it predicts how the scene will change, so 1X can test policies on many situations without building 3D assets. Trained on thousands of hours of EVE humanoid data from homes and offices. 1X released 100+ hours of compressed video under Apache 2.0 plus baseline model weights for a public challenge; its internal production model is not public. In June 2026 1X set up a dedicated World Model Lab.

Used in: EVE

Cloud software launched in March 2024 for deploying and managing fleets of Digit robots. It tracks uptime, throughput and incidents and connects to warehouse and factory management systems through APIs; GXO and Manhattan Associates were early integrations. It is a fleet and workflow platform rather than a published AI model, and Agility says it stays integrated with Digit 5.

Used in: Digit (2023 commercial version), Digit 5

March 2024 model of about 8 billion parameters, trained on internet data and on the physical interaction data gathered by Covariant's warehouse picking robots. It takes text, images, video, joint angles, force readings and suction data, and can output robot commands or predicted video. First aimed at warehouse pick-and-place.

Used in: Covariant Brain warehouse picking robots

Open-source (Apache 2.0) PyTorch library, datasets format and model hub for real-world robot learning. It bundles imitation-learning policies (ACT, Diffusion Policy) and vision-language-action models (π0, SmolVLA and GR00T N1.7), plus drivers for low-cost arms and some humanoids.

Used in: Reachy 2, HopeJR, G1, SO-100 arm, LeKiwi mobile robot

Open 7-billion-parameter vision-language-action model released in June 2024 with its training code; for a while it was the main open baseline that newer models compared against. Tested zero-shot on WidowX and Google robots and fine-tuned on Franka arms.

Used in: WidowX arm (Bridge V2 setup), Franka arm (tabletop and DROID setups)

October 2024 research model built on Gemini that maps raw camera input straight to future driving paths with one network. Waymo presented it as a research paper and listed open problems such as speed and long video. It was not described as running in the robotaxi fleet.

Wayve's end-to-end driving software, learned from data rather than rules and designed to work without HD maps, from driver assistance up to Level 4. It is meant to fit different car computers. Driverless Uber rides using it began in London in September 2026 in a Ford Mustang Mach-E; Wayve also cites a production agreement with Mercedes-Benz.

Used in: Wayve AI Driver robotaxi on Uber (London)

July 2023 model that popularised the term vision-language-action. It took large vision-language models (PaLI-X and PaLM-E) and trained them to output robot actions as well as text, so web knowledge could carry over to robot control. Robot data came from the RT-1 fleet.

Used in: Everyday Robots mobile manipulators (RT-1 data fleet)

Tesla launched FSD (Supervised) v12 in early 2024 using a vision-based, end-to-end neural-network architecture trained on real-world driving data (Tesla, Q1 2024 Update). Its later v14 system uses an end-to-end foundation model trained on customer and Robotaxi data (Tesla, Q4 and FY 2025 Update). Tesla's Robotaxi service launched in Austin in June 2025 (Tesla, 2025 proxy statement).

Used in: Tesla Robotaxi (Model Y service fleet)

In September 2023 TRI showed that 'Diffusion Policy', a generative method borrowed from image AI, could teach robots dexterous skills such as pouring and tool use from a few demonstrations, and framed it as a step towards Large Behavior Models for robots. TRI said it had taught over 60 skills and aimed for 1,000 by the end of 2024. In October 2024 it agreed to work with Boston Dynamics on Atlas.

LINGO-1 (September 2023) commented on driving scenes in plain language and answered questions, without controlling the car. LINGO-2 (April 2024) is a vision-language-action driving model that steers and gives a running explanation, and Wayve says it was the first language-trained driving model tested on public roads, in central London.

Used in: Wayve test car in central London (vehicle not named)

Wayve's generative driving world models. GAIA-1 (2023, later scaled to 9 billion parameters) generates realistic driving video from video, text and actions; GAIA-2 (March 2025) adds multi-camera output and control over weather, lighting, road layout and rare safety-critical events. Wayve described GAIA-4 in August 2026 as powering closed-loop simulation with its AI Driver in the loop.

December 2022 transformer model that reads a short run of camera frames plus a text instruction and outputs robot actions as tokens. It was trained on 130,000 demonstrations collected over 17 months by 13 mobile manipulators. Google released the code; whether the original weights were released is unclear from the sources read.

Used in: Everyday Robots mobile manipulator (13 robots)

Tesla has said since its 2021 AI Day that its humanoid is driven by the same AI approach it builds for its cars, and in mid-2025 the head of Autopilot software also took over the Optimus programme. Tesla has not published a separate named robot model. Elon Musk said in 2026 that the coming AI5 chip is meant to give Optimus enough on-board intelligence to work without a network link.

Used in: Optimus Gen 2