Robot brains

• 1X World ModelWorld model · Commercial

  1X Technologies' generative world model plus VLA stack powering NEO, the consumer-home humanoid shipping via preorder. The world model enables learning by predicting outcomes versus exhaustive pre-programming. Combined with VLA control for action generation. Important verified-vs-claimed flag: NEO home capabilities are teleoperated-assisted; the autonomy-vs-teleoperation distinction is the load-bearing consumer verification point.

• Atlas-GeminiFoundation model · Pilot

  The hybrid brain powering Boston Dynamics' all-electric Atlas. Combines Boston Dynamics' Atlas control stack (including Orbit skill-sharing) with Google DeepMind's Gemini Robotics foundation models for higher-level reasoning and learning. Partnership announced at CES January 2026; Atlas 2026 production is committed to Hyundai plus DeepMind, with the Robotics Metaplant Application Center (RMAC) at Hyundai Metaplant America serving as a supervised-training data factory.

• CarbonFoundation model · Commercial

  Sanctuary AI's Carbon cognitive AI control system. A cognitive hybrid that translates natural language into precise physical actions, with emphasis on human-like hand dexterity. Powers Phoenix, Sanctuary's humanoid known for fine manipulation tasks such as buttoning shirts and handling fragile lab tools. Framed by Sanctuary as 'a brain for work.'

• Gemini RoboticsFoundation model · Research

  Google DeepMind's robotics VLA family. Gemini Robotics builds on Gemini 2.0 with physical actions as an output modality and adds an intermediate reasoning layer for spatial analysis and safety. Gemini Robotics-ER is an embodied-reasoning VLM companion. Gemini Robotics On-Device runs locally, network-independent, and is fine-tunable with 50-100 demonstrations. Gemini Robotics 1.5 introduced 'think before acting'.

• Genie Operator-1 (GO-1)Foundation model · Pilot

  AgiBot's universal embodied foundation model, Genie Operator-1 (GO-1), launched March 2025. Introduces the Vision-Language-Latent-Action (ViLLA) framework: a Vision-Language Model plus a Mixture-of-Experts (a Latent Planner trained on cross-embodiment and human-operation data, and an Action Expert trained on over a million real robot demonstrations). Supports learning from human video, few-shot generalization, cross-embodiment adaptation, and continuous self-evolution.

• GR00T N1 (Isaac GR00T)Foundation model · Research · Open

  NVIDIA's open humanoid foundation model and the first of its kind to be released openly. A dual-system VLA: System 2 uses NVIDIA-Eagle plus SmolLM-1.7B as the VLM at roughly 10 Hz; System 1 is a diffusion transformer producing real-time motor actions. Jointly trained end-to-end on real-robot trajectories, human videos, and synthetic data. Part of the broader Isaac GR00T platform that includes Isaac Lab, Omniverse, and Cosmos.

• GraspVLA (Galbot)Foundation model · Research

  Galbot's grasping foundation model, GraspVLA (launched Jan 2025): an end-to-end Vision-Language-Action model pre-trained on SynGrasp-1B, a billion-frame synthetic grasping dataset with photorealistic rendering and domain randomization. Enables zero-shot generalization to new grasping tasks without additional training, powering dexterous manipulation on Galbot's robots.

• Grok (xAI)Foundation model · Pilot

  xAI's Grok large language model serves as the System 2 conversational and reasoning layer in Tesla Optimus's dual-brain architecture. Grok handles natural-language understanding and high-level instruction reasoning (the "what should I do next" layer), while Tesla's FSD-derived neural networks handle the System 1 visuomotor layer (the "actually doing it" layer). The same Grok-reasoning-plus-Tesla-compute pattern is reused in the "Digital Optimus" (Macrohard) software-agent project announced March 2026.

• Helix (and Helix-02)Foundation model · Commercial

  Figure AI's onboard dual-system VLA. System 2 is an internet-pretrained VLM (7B params, 7-9 Hz) handling scene and language understanding; System 1 is a fast reactive visuomotor policy (80M params, 200 Hz). Helix-02 (January 27, 2026) extended to full-body control via System 0, a 1 kHz neural prior trained on 1,000+ hours of human motion data that replaced roughly 109,504 lines of hand-engineered C++. The first VLA to control a full humanoid upper body including individual fingers from one set of weights, and the first to run two robots from one set of weights.

• Mentee BrainFoundation model · Research

  Mentee Robotics' end-to-end neural-network AI for the MenteeBot humanoid (Mentee Brain): integrates perception, reasoning, and motor control, trained via a proprietary Sim2Real pipeline, with NeRF-based on-the-fly environment mapping and transformer-based LLMs for interpreting commands and planning tasks. An AI-first stack spanning all operational layers. Mentee Robotics was acquired by Mobileye (Jan 2026) to lead its Physical AI division.

• Mi-Sense (Xiaomi)Research model · Research

  Xiaomi's perception and interaction AI for the CyberOne humanoid (unveiled 2022): the Mi-Sense depth-vision module for 3D spatial perception and the MiAI engine for vocal-emotion and environmental-sound recognition. Recognizes 45 human-emotion classifications and 85 environmental sounds, with real-time facial-expression and gesture detection. A research/showcase-stage system rather than a generalist control foundation model.

• Neuraverse Cognitive StackOS-layer · Pilot

  NEURA Robotics' hybrid neural plus symbolic multimodal cognitive stack. Neuraverse is a shared OS and learning platform connecting robots so that one robot's learned skill propagates to others. Positioned by NEURA as 'invisible OS for the World of Things.' At CES 2026, NEURA's robots are powered by NVIDIA Isaac GR00T XX, with Neuraverse acting as the orchestration layer atop GR00T.

• Redwood AIFoundation model · Pilot

  Redwood AI is 1X Technologies' onboard vision-language-action control model for the NEO humanoid, recorded in the registry as a brain distinct from the separate 1X World Model. A roughly 160-million-parameter vision-language transformer tailored to the humanoid form factor, Redwood runs fully onboard NEO's embedded GPU at about 5 Hz and serves as the deployed real-time control and action policy, performing end-to-end mobile bi-manual manipulation, whole-body control, navigation, retrieving objects, and opening doors, trained on teleoperated and autonomous EVE and NEO episodes including learning from failure rollouts. It is distinct from the 1X World Model, a generative physics-grounded world model released in January 2026 for prediction, training, and evaluation: Redwood is the on-robot controller while the World Model is a generative simulation and learning substrate whose outputs are translated to motion by a separate inverse-dynamics model, so the two differ in model class, deployment role, release date, and training method rather than one being a rebrand of the other. Released around June 10, 2025, Redwood is closed and proprietary and is recorded at pilot maturity: it runs on real early-access NEO units offered at twenty thousand dollars or four hundred ninety-nine dollars per month, but 1X describes it as early in development that does not always succeed on the first try, and autonomous operation in customer homes is a stated goal rather than verified, so it is not presented as mass-shipped production.

• RobotEra ERA-42Foundation model · Pilot

  RobotEra's end-to-end Vision-Language-Action model, ERA-42 (introduced late 2024), described as the world's first truly embodied large model for five-fingered robot control. Integrates vision, language, and proprioceptive data in a unified on-device neural network for high-frequency, real-time motor control without cloud assistance; powers the RobotEra L7 humanoid. Developed under founder Chen Jianyu (Tsinghua University).

• Tesla FSD-BotFoundation model · Pilot

  Tesla's neural-net brain derived from the FSD vehicle stack plus Dojo training infrastructure. V3 is expected to run on Tesla's AI5 inference chip (taped out April 2026); Musk states AI5 has roughly 5x the memory bandwidth of its predecessor (AI4) and will ship first in Optimus, though these remain Musk-stated figures pending independent verification. On-device, vision-based, sharing architecture with Tesla vehicles. The Tesla AI-silicon advantage (FSD plus Dojo) is a real engineering edge. Less specified than Figure 03's stack was at the equivalent stage; full V3 specs are unknown as of mid-2026, with a summer 2026 unveil expected. Verified-vs-claimed: FSD-as-precedent is mixed; 'almost done' for nearly a decade.

• UBTech BrainNet (Thinker)Foundation model · Pilot

  UBTech's embodied-AI system for the Walker S2 industrial humanoid: the BrainNet 2.0 platform (multimodal reasoning, environmental understanding, and high-level task orchestration over stereo vision, force-feedback joints, and IMUs), combined with the Thinker embodied-AI foundation model, the Thinker-WM world model, and the Co-Agent multi-robot system, forming a dual-loop architecture for production-line swarm intelligence.

• UnifoLMWorld model · Research · Open

  Unitree Robotics' open-source unified-large-model series for general-purpose robot learning: comprising UnifoLM-WMA-0 (world-model-action) and UnifoLM-VLA-0 (vision-language-action). Unitree's G1 humanoid integrates UnifoLM-WMA-0.

• VLT (XPENG VLA / VLA 2.0)Foundation model · Pilot · Open

  Xpeng Robotics' 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. Announced for open-sourcing to global partners, with Volkswagen as the launch partner. VLA 2.0 rolled to Xpeng Ultra vehicles in Q1 2026; IRON humanoid mass production is targeted for end-2026.

Machine-readable: this page as markdown.