AI stacks compared: what actually drives each robot
The AI brain that drives a robot is a different thing from its body, and the stacks diverge sharply. DEPLOY tracks 59 verified brain records. Some are captive: Helix is a proprietary dual-system vision-language-action model that runs only Figure's own robots, and 1X's Redwood and Tesla's FSD-derived stack follow the same vertically integrated pattern. Others are open and licensable: NVIDIA GR00T and Physical Intelligence's pi0 are foundation models other makers can build on. Many driver stacks (Waymo, Zoox, Nuro, WeRide) keep their architecture undisclosed, which DEPLOY records as an honest gap rather than guessing.
Brain records are sourced from the DEPLOY registry. How we verify.
59
Verified brain records
45
Powering at least one robot
8
Open source
45
Foundation models and VLAs
Captive, third-party, or research-only
Beyond what kind of model a brain is, the sharpest axis is who can use it. DEPLOY's framework names four provider architectures:
- Captive (vertically integrated): Brain entity exists under the same corporate state as the hardware it powers. Not licensable to third parties; vertically integrated. The brain entity's roadmap, capabilities, and commercial trajectory are inseparable from the hardware roadmap.
- Third-party foundation model: Brain entity licensed or deployed across multiple unaffiliated hardware partners. Foundation-model-tier capability scope (cross-domain manipulation, cross-platform deployment). The brain entity's roadmap is partner-agnostic; commercial trajectory tied to broad adoption rather than a single hardware customer.
- Hybrid (captive + third-party): Brain entity has both captive-deployment AND third-party-licensing channels. The brain powers a parent's hardware AND is licensable to unaffiliated partners. Distinct from pure-captive (no licensing) and pure-third-party (no captive hardware).
- Research-only: Brain entity in development or laboratory stage; no verified commercial deployment. May target captive or third-party future architecture but the framework records the current state as research, not commercial.
Undisclosed is not unknown-away
When a maker does not publish how its brain is built, DEPLOY records the architecture as not disclosed rather than inventing one. That gap is itself a finding: the most secretive stacks are often the most commercially advanced driver systems, where the architecture is the competitive moat.
Foundation models and VLAs
Vision-language-action and end-to-end driving models: the generalist stacks that translate perception into motor action.
| Brain | Architecture | Access | Developer | Powers these robots | DEPLOY maturity |
|---|---|---|---|---|---|
| Not disclosed | Proprietary | Aurora Innovation | Aurora Driver on Peterbilt 579 (PACCAR), Aurora Driver on International LT (2nd-gen), Volvo VNL Autonomous, Aurora Driver | Commercial | |
Tesla FSD-BotCaptive (vertically integrated) | Neural-net brain derived from the Tesla FSD stack plus Dojo training. V3 runs on the Tesla AI5 chip (roughly 5x memory bandwidth of predecessor). On-device, vision-based, shares architecture with Tesla vehicles. | Proprietary | Tesla | Tesla Optimus, Tesla Robotaxi (Model Y ADS), Cybercab, Semi | Pilot |
| Not disclosed | Proprietary | Nuro | Nuro R2, Nuro R3, Nuro Autonomous Prius | Pilot | |
| Not disclosed | Proprietary | Skydio | Skydio X10, Skydio X10D, Skydio Dock | Production | |
| Not disclosed | Proprietary | Avride | Avride Robotaxi (Hyundai Ioniq 5), Avride delivery robot | Pilot | |
| Not disclosed | Proprietary | Embark Trucks | Embark Autonomous Truck, Embark Driver | Research | |
| Not disclosed | Proprietary | Gatik | Gatik Autonomous Box Truck, Gatik Autonomous Drive | Commercial | |
| Not disclosed | Proprietary | AgiBot | AgiBot Yuanzheng A2, AgiBot Lingxi X2 | Pilot | |
| Dual-system VLA. System 2: VLM (NVIDIA-Eagle + SmolLM-1.7B) at roughly 10 Hz. System 1: diffusion transformer producing real-time motor actions. Jointly trained end-to-end on real-robot trajectories, human videos, and synthetic data. Runs on Jetson AGX Thor at the edge. | Open source | NVIDIA | AEON, Digit | Research | |
| Dual-system VLA. System 2 (S2): onboard internet-pretrained VLM, 7B params, 7-9 Hz. System 1 (S1): fast reactive visuomotor policy, 80M params, 200 Hz. Helix-02 added System 0, a 1 kHz neural prior trained on 1,000+ hours of human motion data, replacing roughly 109,504 lines of hand-engineered C++. Runs on embedded low-power onboard GPUs, no cloud. | Proprietary | Figure AI | Figure 02, Figure 03 | Commercial | |
| Not disclosed | Proprietary | Kodiak AI | Kodiak RoboTruck, Kodiak Driver | Commercial | |
| Not disclosed | Proprietary | Plus | Plus SuperDrive Autonomous Truck, PlusDrive | Pilot | |
| Not disclosed | Proprietary | Shield AI | Shield AI V-BAT, Anduril Fury (YFQ-44A) | Production | |
| Not disclosed | Proprietary | UBTech Robotics | Walker S2, Walker S Lite | Pilot | |
| Vision-centric VLA (Tesla-FSD-style, vision-only), shared across Xpeng EVs, robotaxis, and the IRON humanoid. Trained on a 30,000+ GPU cloud cluster. Runs on the Xpeng Turing AI chip. | Open source | XPeng Robotics | XPeng IRON, XPENG Robotaxi | Pilot | |
| Not disclosed | Proprietary | Waabi | Volvo VNL Autonomous, Waabi Driver Autonomous Truck | Pilot | |
| Not disclosed | Proprietary | Waymo | Waymo Driver (6th gen), Waymo Via (autonomous trucking) | Commercial | |
| Not disclosed | Proprietary | Zipline | Zipline Platform 1 (Zip), Zipline Platform 2 (P2) | Production | |
| Not disclosed | Proprietary | Baidu | Apollo RT6 | Commercial | |
| Hybrid stack: Boston Dynamics Atlas control software (including Orbit skill-sharing) integrated with Google DeepMind Gemini Robotics foundation models for higher-level reasoning and learning. | Proprietary | Boston Dynamics, Google DeepMind | Atlas | Pilot | |
| Not disclosed | Proprietary | Bot Auto | Bot Auto Autonomous Truck | Commercial | |
| Cognitive hybrid translating natural language into precise physical actions. Emphasis on human-like hand dexterity for fine-manipulation tasks. | Proprietary | Sanctuary AI | Phoenix | Commercial | |
| Not disclosed | Proprietary | Einride | Einride Pod | Commercial | |
| VLA built on Gemini 2.0 with physical actions as output modality, plus an intermediate reasoning layer for spatial analysis and safety. Gemini Robotics-ER is the embodied-reasoning VLM companion. Gemini Robotics On-Device runs locally, network-independent, and is fine-tunable with 50-100 demonstrations. Gemini Robotics 1.5 emphasizes a 'think before acting' chain-of-thought style. | Proprietary | Google DeepMind | Apollo | Research | |
| Not disclosed | Proprietary | Galbot | Galbot G1 | Research | |
| Large language model (System 2 reasoning) deployed as the conversational and high-level instruction layer of Tesla Optimus's dual-brain stack. Paired with Tesla's FSD-derived visuomotor neural networks (System 1, see /brains/tesla-fsd-bot). | Proprietary | xAI | Tesla Optimus | Pilot | |
| Not disclosed | Proprietary | IntBot | José (IntBot humanoid) | Pilot | |
| Not disclosed | Proprietary | Mentee Robotics | MenteeBot | Research | |
Redwood AICaptive (vertically integrated) | 160M-parameter vision-language-action (VLA) transformer; onboard, ~5Hz; end-to-end mobile manipulation | Proprietary | 1X Technologies | NEO | Pilot |
| Not disclosed | Proprietary | Robot Era | RobotEra L7 | Pilot | |
| Not disclosed | Proprietary | Stack AV | Stack AV Autonomous Truck (Peterbilt 579) | Research | |
| Not disclosed | Proprietary | Torc Robotics | Autonomous Freightliner Cascadia (Torc) | Pilot | |
| Not disclosed | Proprietary | TuSimple | TuSimple Autonomous Truck | Research | |
| Not disclosed | Proprietary | WeRide | Robotaxi GXR | Commercial | |
| Not disclosed | Proprietary | Zoox | Zoox Robotaxi | Pilot | |
| Dexterous robot foundation model (VLA) for sustained autonomous operation on a pair of stationary dual robotic arms (the Dyna Robotics 'Dynasaur' system, not humanoid). | Proprietary | Dyna Robotics | No registry robot yet | Commercial | |
| Physics-first, embodiment-agnostic field/mobility autonomy foundation models for GPS-/map-denied unstructured environments (navigation + mobile autonomy without maps/GPS/pre-defined trajectories). | Proprietary | Field AI | No registry robot yet | Commercial | |
| Embodied/manipulation foundation model (generalist VLA-class) trained on raw physical-interaction data; ~7B-param scaling phase transition reported. | Proprietary | Generalist AI | No registry robot yet | Research | |
| VLA flow model, roughly 3B-parameter VLM backbone, motor commands at up to 50 Hz, cross-embodiment across 8 platforms: UR5e, Bimanual UR5e, Franka, Bimanual Trossen, Bimanual ARX, Mobile Trossen, Mobile Fibocom. Trained on internet-scale vision-language data, Open X-Embodiment, and Physical Intelligence's dexterous-manipulation dataset. PyTorch and JAX implementations. | Open source | Physical Intelligence | No registry robot yet | Research | |
| Vision-language-action model built on pi0; co-trained on heterogeneous multi-robot + web data with high-level semantic subtask prediction for open-world generalization. | Open source | Physical Intelligence | No registry robot yet | Research | |
| Multimodal any-to-any autoregressive sequence model (~8B params); next-token prediction over text/images/video/robot-actions/sensors. | Proprietary | Covariant | No registry robot yet | Research | |
| Robot foundation model; specific architectural details are limited in public disclosure this pass and should be firmed at next pass. | Proprietary | RoboForce | No registry robot yet | Pilot | |
| 8B parameter VLM initialized from vision-language model for grounding (pointing, counting, object localization). Navigation via pointing (infers image coordinates of target location) with fallback to metric displacements. Trained with prefix-caching (tree-based attention masking, 22x token reduction) and online RL (CISPO algorithm, +3.2% improvement). | Proprietary | Mistral AI | No registry robot yet | Research | |
| Unified 'omni-bodied' foundation model controlling any robot form without prior body-form knowledge (quadrupeds, humanoids, tabletop arms, mobile manipulators). Trained on online human videos plus physics simulations across thousands of form factors. Built-in force-limiting safety constraints. | Proprietary | Skild AI | No registry robot yet | Commercial | |
| End-to-end embodied driving foundation model (AV2.0); paired with GAIA latent-diffusion generative world models (sim/training) + LINGO VLA interpretability layer. | Proprietary | Wayve | No registry robot yet | Pilot |
World models
Models that predict how the physical world evolves, used for planning, simulation, and policy learning.
| Brain | Architecture | Access | Developer | Powers these robots | DEPLOY maturity |
|---|---|---|---|---|---|
| Unitree's open-source unified-large-model series for general-purpose robot learning across multiple embodiments. UnifoLM-WMA-0 (world-model–action): a world model of robot–environment physical interaction, used both as a simulation engine for synthetic data and as a policy-enhancement head that predicts future environmental states (video prediction translated into limb actions). UnifoLM-VLA-0 (vision–language–action): a VLA for general-purpose humanoid manipulation, built on a UnifoLM-VLM base. Open-sourced on Unitree's GitHub (WMA-0 Sept 2025; VLA-0 Jan 2026). | Open source | Unitree Robotics | Unitree G1, Unitree R1, Unitree H1, Unitree H2 | Research | |
| Generative world model enabling learning by predicting outcomes versus exhaustive pre-programming; combined with VLA control for action generation. | Proprietary | 1X Technologies | NEO, EVE | Commercial | |
| Not disclosed | Proprietary | Pony AI | Pony Gen-7 Robotaxi | Commercial | |
| Self-supervised joint-embedding predictive video world model (~1.2B params); action-conditioned V-JEPA 2-AC variant for robot control. | Open source | Meta | No registry robot yet | Research |
Frameworks
Open robot-learning frameworks that host and train other brains rather than drive a robot directly.
| Brain | Architecture | Access | Developer | Powers these robots | DEPLOY maturity |
|---|---|---|---|---|---|
| Not disclosed | Proprietary | AeroVironment | AeroVironment Puma, AeroVironment P550 | Production | |
| Open-source robot-learning framework. Library-style hosting of VLA models (pi0, pi0.5, GR00T N1.5) plus datasets, simulation environments (LIBERO, Meta-World), training tooling, multi-GPU support, plugin system. Released v0.4.0 with the upgrades above. | Open source | Hugging Face | No registry robot yet | Production |
OS-layers and autonomy stacks
Autonomy operating systems licensed to hardware makers as the driving layer of their machines.
| Brain | Architecture | Access | Developer | Powers these robots | DEPLOY maturity |
|---|---|---|---|---|---|
| Not disclosed | Proprietary | Anduril Industries | Anduril Roadrunner, Anduril Bolt, Anduril Fury (YFQ-44A), Ghost (Ghost-X), Dive-LD, Ghost Shark (XL-AUV) | Production | |
| Cloud-connected autonomy OS: teach-and-repeat / mapped autonomous routing with onboard sensor-based obstacle avoidance, plus a cloud portal and mobile app for fleet management and reporting. Licensed to OEMs as the autonomy layer of their cleaning hardware. | Proprietary | Brain Corp | SoftBank Whiz, Tennant X4 ROVR, ICE Cobi 18, Nilfisk Liberty SC60, Tennant T7AMR | Production | |
| Not disclosed | Proprietary | Mobileye | Volvo EX90, Mobileye Drive, VW ID. Buzz AD | Pilot | |
| Hybrid neural and symbolic multimodal cognitive stack. Plus Neuraverse, a shared OS and learning platform that connects robots so that one robot's learned skill propagates to others. NEURA's CES 2026 robots are powered by NVIDIA Isaac GR00T XX; Neuraverse is the orchestration layer atop GR00T. | Proprietary | NEURA Robotics | 4NE-1, MiPA | Pilot | |
| Not disclosed | Proprietary | Wing | Wing Delivery Aircraft | Production |
Research models
Research-stage brains not yet carried into a commercial product in the registry.
| Brain | Architecture | Access | Developer | Powers these robots | DEPLOY maturity |
|---|---|---|---|---|---|
| Not disclosed | Proprietary | Xiaomi | Xiaomi CyberOne | Research | |
| 7B-parameter VLA. Architecture and training-data details pending firmer primary-source confirmation (OpenVLA paper at openvla.github.io and arXiv). | Open source | Google DeepMind | No registry robot yet | Research | |
| Transformer VLA trained on web text and images, directly outputs robot actions. Instantiations on PaLM-E and PaLI-X. RT-2-X is 55B parameters. Chain-of-thought reasoning for long-horizon planning. Trained on web data plus Open X-Embodiment. | Proprietary | Google DeepMind | No registry robot yet | Research |
Get one email when a robot's AI stack changes.
Browse the registryAll verified answers
Brain records are drawn from the DEPLOY registry: reviewed Brain entities with their developer companies and the models they power. Architecture is the record's own disclosed detail, or marked not disclosed where the maker has not published it. Maturity is DEPLOY's verified stage. The captive vs third-party provider-architecture label renders only where DEPLOY has editorially classified the brain, and states honest absence otherwise. How we verify
Related answers
Robot brains
Every AI brain in the registry: foundation models, world models, frameworks, and OS-layers.
Read article →
Humanoid robot autonomy: verified vs claimed
What makers claim about autonomy set against what DEPLOY has verified.
Read article →
All humanoid robots
Every humanoid in the registry, with maturity, deployments, and specs.
Read article →
Compare
All robot brainsAutonomy: verified vs claimedAll humanoid robots
Brain developers in the registry
- 1X Technologies
- AeroVironment
- AgiBot
- Anduril Industries
- Aurora Innovation
- Avride
- Baidu
- Boston Dynamics
- Bot Auto
- Brain Corp
- Covariant
- Dyna Robotics
- Einride
- Embark Trucks
- Field AI
- Figure AI
- Galbot
- Gatik
- Generalist AI
- Google DeepMind
- Hugging Face
- IntBot
- Kodiak AI
- Mentee Robotics
- Meta
- Mistral AI
- Mobileye
- NEURA Robotics
- Nuro
- NVIDIA
- Physical Intelligence
- Plus
- Pony AI
- RoboForce
- Robot Era
- Sanctuary AI
- Shield AI
- Skild AI
- Skydio
- Stack AV
- Tesla
- Torc Robotics
- TuSimple
- UBTech Robotics
- Unitree Robotics
- Waabi
- Waymo
- Wayve
- WeRide
- Wing
- xAI
- Xiaomi
- XPeng Robotics
- Zipline
- Zoox
Part ofRobot brains/Humanoid robots/Robotaxis
See also: all robot brains, autonomy verified vs claimed, humanoid robots.
For AI assistants
Use DEPLOY in Claude, Cursor, and other AI assistants
Connect any MCP-aware assistant to the live registry. Ask about robot deployments, incidents, and regulations and get answers grounded in verified data instead of training memory.
API: GET /v1/entities?type=brain · canonical URL: /ai-stacks · machine-readable mirror: /ai-stacks.md
Want your own assistant to answer like this? Plug DEPLOY into ChatGPT or Claude so it checks the real facts before it answers.