Models & Training
The AI substrate underneath physical robots: world models, simulation and RL training infrastructure, trained models such as manipulation and affective/social-interaction models, and the datasets — robot manipulation, egocentric video, gameplay — used to train them.
RI Dataset Best Practices: a real, independently verifiable GitHub repository — or, for well-established datasets a direct link to the authors' own original release; licenses reported exactly as found; and every entry checked against primary sources — the paper, the official repo or release, the project's own site.
World Models
Simulation & Training
Models
Datasets
Sort: Default
Sort: Most stars
Sort: Recently updated
The Models & Training catalog couldn't be loaded right now. Try refreshing the page in a moment.
Live GitHub stats (stars, license, last update) aren't available right now. Project links below still work — this usually clears up on its own shortly.
World Models
Open Verified against source
Closed Vendor-described, unverified
No world models match your search.
Simulation & Training
Physics Engine
RL Training
Synthetic Data & Sensors
No simulation or training tools match your search.
Models
Manipulation Models
Affective & Social Interaction
No models match your search.
Datasets
Robot Manipulation
Egocentric Human Video
Gameplay / Simulated Interaction
Human Motion Capture
No datasets match your search.
World Models
Open
NVIDIA Cosmos — World foundation model platform with a dedicated action-conditioned mode that predicts forward/inverse dynamics for robotics, cameMeta V-JEPA 2 — Self-supervised video world model; the V-JEPA 2-AC variant solves robot manipulation tasks — reaching, grasping, pick-and-place — Oasis — Interactive world model from Decart and Etched using a diffusion transformer to generate real-time, playable gameplay frame-by-fraDrivingWorld — Video-GPT-style world model for autonomous driving, generating high-fidelity, controllable future driving video and ego-state predDreamerV3 — General world-model-based RL algorithm that masters diverse domains — Atari, DMLab, Minecraft diamond collection, real robot locomTD-MPC2 — Model-based RL algorithm pairing a learned latent world model with local trajectory optimization (MPC); a single set of hyperparamSkywork Matrix-Game — Interactive world model generating real-time, controllable game-style video (720p at 40 FPS with the 5B-parameter model) with minu
Closed
Google DeepMind Genie 3 — Real-time interactive world model generating explorable environments from a single text or image prompt. Described from DeepMind'sWorld Labs Marble — World Labs' first commercial world model, generating persistent, explorable 3D environments from text, image, or video input. DescWayve GAIA-3 — Wayve's generative world model for autonomous driving, now at GAIA-3 (15B parameters, evolved from the original 2023 GAIA-1) — use1X World Model — 1X's world model for its EVE/NEO humanoids, trained on real robot data to predict how environments evolve in response to actions —Runway GWM Robotics — The robotics variant of Runway's GWM-1 general world model: an autoregressive video-generation model simulating robot manipulat
Simulation & Training
Physics Engine
MuJoCo — Contact/soft-body physics engine maintained by Google DeepMind — the reference benchmark for simulation accuracy that newer GPU enGenesis — GPU-accelerated multi-physics engine (rigid body, FEM, MPM, particle-based) with a photorealistic renderer, purpose-built for roboBullet — Long-running open-source collision detection and rigid/soft-body physics engine; the basis for PyBullet and used inside Blender anNVIDIA Isaac Lab — GPU-accelerated robot learning framework built on Isaac Sim, for large-scale parallel reinforcement learning and imitation learninProject Chrono — Multi-physics simulation engine covering rigid/flexible bodies, granular dynamics, and fluid-structure interaction.Brax — Differentiable JAX-based physics engine for massively parallel simulation on accelerator hardware. Worth flagging: the project's oDrake — Model-based design and verification toolkit from MIT/Toyota Research Institute — C++ dynamical-systems modeling, numerical integraNewton Physics — GPU-accelerated physics simulation engine built on NVIDIA Warp, developed jointly by NVIDIA, Google DeepMind, and Disney Research MuJoCo Playground — Google DeepMind's open-source library of GPU-accelerated MuJoCo (MJX) environments for robot learning and sim-to-real transfer, spGazebo (gz-sim) — The current generation of Gazebo (formerly "Ignition Gazebo"), the long-running open-source robotics simulator providing physics, Webots — Cyberbotics' robot simulator, open-sourced under Apache-2.0 in 2018 after two decades as a commercial product; includes a full phy
RL Training
Stable-Baselines3 — Reliable PyTorch implementations of core reinforcement learning algorithms (PPO, SAC, TD3, DQN) with a consistent, well-documentedCleanRL — Single-file, research-friendly reference implementations of major RL algorithms, built for readability and reproducibility rather Ray RLlib — Distributed reinforcement learning library within the Ray project, used for training at a scale beyond a single machine.TorchRL — PyTorch-native reinforcement learning toolkit built around the TensorDict primitive, which passes structured data across environmeLeRobot — Hugging Face's end-to-end robot learning library — imitation learning, RL training, and dataset tooling in one framework — used toRSL-RL — ETH Zurich Robotic Systems Lab's fast, minimal PPO-based RL training library, widely used to train legged-robot locomotion policie
Synthetic Data & Sensors
NVIDIA Isaac Sim — Full robot simulation platform on Omniverse with RTX and physics-based sensor models and a built-in synthetic data generation toolManiSkill — GPU-parallelized robot simulation and manipulation benchmark built on SAPIEN, capable of 30,000+ FPS synthetic visual data collectNiantic Spatial Places Library (real-to-sim scenes) — Real-world captured warehouse and street environments delivered as simulation-ready USDZ (Gaussian splat plus aligned collision meHabitat-Sim — Meta AI's high-performance 3D simulator for embodied AI research, rendering photorealistic indoor scenes at thousands of frames peEidon Sim — Open-source simulation environment and visualizer from Eidon AI, released when the company shut down: quaternion-to-Euler conversion, a seven-DOF arm kinematics solver, 3D skeletal reconstruction, and trajectory playback synced to video, with simulation-compatible output. Built around motion captured by Eidon's own wearable trackers.OmniGibson (BEHAVIOR-1K) — Stanford's simulation platform built on NVIDIA Omniverse/Isaac Sim, now published under the BEHAVIOR-1K project name, targeting 1,robosuite — Modular MuJoCo-based simulation framework and benchmark suite for robot manipulation research from the ARISE Initiative (Stanford/
Models
Manipulation Models
NVIDIA GraspGenX — Cross-embodiment 6-DOF grasp generation model: one network conditioned on a gripper's swept volume, generalizing to grippers unseeNVIDIA Isaac GR00T (N1.7) — NVIDIA's open humanoid/cross-embodiment VLA foundation model; N1.7 adds 20K hours of EgoScale human video to pretraining. Pairs wiOcto — Berkeley's open generalist transformer policy trained on Open X-Embodiment; designed to be fine-tuned to new robots, sensors, and RDT-1B (Robotics Diffusion Transformer) — Tsinghua's 1B-parameter diffusion foundation model for bimanual manipulation.Diffusion Policy — Columbia/Stanford's visuomotor policy that generates robot actions via denoising diffusion; a widely used baseline for imitation lACT (Action Chunking with Transformers) — Imitation-learning policy behind ALOHA bimanual manipulation; predicts chunks of future actions for smooth fine-grained control.OpenVLA — 7B-parameter open vision-language-action model from Stanford, UC Berkeley, and Toyota Research Institute, trained on 970K trajectoropenpi (π0 / π0-FAST / π0.5) — Physical Intelligence's open release of its π0 flow-based VLA, the π0-FAST autoregressive variant, and π0.5 — base checkpoiRT-2 — Google DeepMind's vision-language-action model, casting robot actions as text tokens co-trained with internet-scale vision-languGemini Robotics (ER 2 / Robotics 2) — Google DeepMind's robotics model family: Gemini Robotics-ER 2, an embodied-reasoning vision-language model, is available via the1X Redwood — 1X's onboard vision-language-action model for its NEO Gamma humanoid: a 160M-parameter transformer decoding actions via diffusFigure Helix (Helix 02 / Helix 2.5) — Figure's dual-system VLA for its humanoids: a 7B-parameter System 2 vision-language model paired with an 80M-parameter System 1π0.7 (Physical Intelligence) — Physical Intelligence's successor to its open π0/π0.5 models, trained with context conditioning on heterogeneous data to make it
Affective & Social Interaction
SenseVoice (FunAudioLLM) — Speech recognition model that also tags emotion and non-speech audio events (laughter, applause) in the same pass — the perceptionCosyVoice (FunAudioLLM) — Text-to-speech model accepting natural-language instructions for emotion, style, and speaking rate — the expressive half of the piEmotiVoice (NetEase Youdao) — Multi-voice, prompt-controlled emotional TTS engine with 2,000+ voices and explicit emotion tags — a fully commercially-licensed aMoshi (Kyutai) — Full-duplex spoken-dialogue foundation model that listens and speaks simultaneously rather than turn-by-turn, built on the Mimi stRetico — Incremental spoken-dialogue framework for chaining perception, reasoning, and generation modules with sub-utterance streaming lateEmpatheticDialogues (Meta) — 25k crowd-sourced conversations grounded in a speaker's described emotional situation, tagged with one of 32 emotion labels — a wiESConv (Tsinghua) — Emotional-support conversation dataset with annotated support strategies (e.g. reflection, reframing) exchanged between a help-seeMELD — Multimodal (text, audio, video) emotion-labeled dialogue dataset built from Friends TV series clips — used for multi-party emotion
Datasets
Robot Manipulation
Open X-Embodiment — Google DeepMind's unified cross-embodiment robot dataset, consolidating dozens of separately collected manipulation datasets into BridgeData V2 — 60,096 real-world manipulation trajectories across 24 environments and 13 skills, collected on a low-cost WidowX 250 arm — built tLanguage-Table — Human-collected robot manipulation dataset pairing natural-language instructions with tabletop block-rearrangement demonstrations,RoboNet — ~162,000 manipulation trajectories (15 million frames) collected across 7 different robot platforms (Sawyer, Baxter, WidowX, FrankABC-130K — 130,703 teleoperated bimanual manipulation episodes (~3,590 hours) across 195 tasks, collected on dual-arm 6-DoF robot stations byMobile ALOHA — Stanford's mobile-base bimanual teleoperation platform and household manipulation demonstrations (cooking, cleaning, furniture tasMimicGen — NVIDIA and UT Austin's data-generation system that synthesizes large demonstration datasets from a small number of human demos viaDexMimicGen — Extends MimicGen's automated data-generation approach to bimanual dexterous-hand manipulation tasks.RoboCasa / RoboCasa365 — Large-scale simulation of everyday household tasks for generalist robots; RoboCasa365 expands the original 120-kitchen-scene releaEgoMimic — Georgia Tech's method and dataset pairing Project Aria egocentric human video with matched bimanual robot teleoperation demonstratGRAIL — NVIDIA's synthetic data-generation pipeline producing physics-validated 4D human-object-interaction trajectories retargeted to theFurnitureBench — KAIST CLVR lab's long-horizon furniture-assembly manipulation benchmark on a real Franka arm (plus an Isaac-Gym simulated counterpManipArena — A real-robot bimanual manipulation benchmark (20 tasks spanning execution reasoning, semantic reasoning, and mobile manipulation) UniDex — A robot foundation suite for universal dexterous-hand control, including 50,000+ trajectories across 8 dexterous hands retargeted MimicPlay — Stanford and NVIDIA's hierarchical imitation-learning framework that trains a high-level planner on cheap unlabeled human "play" dUniversal Manipulation Interface (UMI) — A portable, robot-agnostic handheld gripper and data pipeline that turns in-the-wild human manipulation demonstrations into robot robomimic — Stanford ARISE Initiative's framework and datasets (proficient-human, multi-human, and machine-generated demonstrations) for studyCALVIN — University of Freiburg's benchmark for language-conditioned long-horizon manipulation — 24 hours of simulated teleoperated "play" ManiSkill2 — SAPIEN-powered simulated manipulation benchmark with 4,000,000+ demonstration frames across 20 skill families and 2,000+ objects; RoboArena — A distributed, real-world evaluation system for generalist robot policies built on a standardized DROID-based platform (Franka armLESS (Local Encoder for Spatial Sensing) — Technion's tactile-imaging dataset — 800+ hours of poke and palpation trajectories collected with a Franka arm and a gel-based tacRoboDojo — A unified sim-and-real benchmark for generalist robot manipulation policies, spanning 42 Isaac Sim tasks and 18 real-world tasks aFunctional Manipulation Benchmark (FMB) — UC Berkeley RAIL's benchmark for diverse, contact-rich manipulation — 22,550 demonstrations of single- and multi-stage assembly taRoboCOIN — BAAI/FlagOpen's open bimanual manipulation dataset spanning 15 robot platforms, 421 tasks, and 180,000+ demonstrations, with roughOpenHLM-data — Training data for whole-body humanoid loco-manipulation on the Unitree G1, combining teleoperation and human-motion-capture pipeliΨ0 / psi-data — Training data (real-world task demos, simulated data, pretraining data) behind Ψ0, an open foundation model for universal humanoidHumanoid Everyday — Unitree G1/H1 demonstrations across roughly 260 everyday household tasks (RGB, depth, and LiDAR at 30Hz, ~500GB); raw data hosted RT-1 — Google DeepMind's foundational robotics transformer dataset — roughly 130,000 real-robot episodes across 700+ tasks collected on 1MolmoAct Dataset — Ai2's real-world manipulation dataset behind MolmoAct — 10,689 trajectories on a single-arm Franka across 93 tasks (7,730 home traMolmoAct2 Bimanual YAM Dataset — Ai2's bimanual manipulation dataset for MolmoAct2 — roughly 34,500 demonstrations (720+ hours) across 28+ tasks on a bimanual YAM MolmoAct2 SO-100/101 Dataset — Ai2's community-sourced manipulation dataset for MolmoAct2 — 38,059 episodes (19.8M frames, ~184 hours) curated and quality-gated RoboMIND 2.0 — A large Chinese cross-embodiment manipulation dataset — 310,000+ bimanual trajectories across 6 embodiments (including a TienKung NVIDIA GR00T X-Embodiment Sim — NVIDIA's simulated post-training dataset for GR00T N1 — roughly 425,000 trajectories (1.91TB) spanning cross-embodiment bimanual mRoboTurk — Stanford's crowdsourced teleoperation platform and dataset — remote workers control a real Sawyer robot arm through a web/gamifiedDROID — Distributed Robot Interaction Dataset — 76,000 real-robot manipulation trajectories (350 hours) collected across 564 scenes and 86AgiBot World — Large-scale bimanual manipulation dataset from AgiBot (Zhiyuan Robotics) — IROS 2025 Best Paper Award Finalist — 1,003,672 traject
Egocentric Human Video
EgoVerse — A growing egocentric human demonstration dataset for robot learning (1,362 hours, 80,000 episodes across 1,965 tasks), aggregatingEgoHumanoid (sample) — OpenDriveLab's sample release pairing egocentric human demonstrations (PICO VR, ZED Mini) with a smaller Unitree G1 teleoperation PH2D — "Physical Human-Humanoid Data" — egocentric human task demonstrations (Apple Vision Pro, Meta Quest 3) paired with real and simulaTouch and Go — ~13,900 human touches (246,000 paired vision-tactile frames) across ~3,971 real-world objects, captured in the wild with a GelSigh
Gameplay / Simulated Interaction
OpenAI VPT (Video PreTraining) — Contractor Minecraft gameplay recordings — video paired with action logs — that OpenAI used to train agents via behavioral cloningTEACh — Amazon's dataset of human-human interactive dialogues collected to complete household tasks inside the AI2-THOR simulator — one peALFRED — A benchmark pairing natural-language instructions and egocentric vision with expert action sequences for long-horizon household ta
Human Motion Capture
CORE4D — 4D human-human-object interaction dataset — two people collaboratively rearranging objects, captured via motion capture for futureMOSAIC — BAAI's large motion bank for humanoid whole-body motion tracking and teleoperation, blending optical and inertial mocap with generPHUMA — KAIST's physics-filtered humanoid locomotion dataset, retargeting large-scale human motion capture to Unitree G1/H1-2 while correcG1 Moves — 60 motion-capture-to-Unitree-G1-retargeted clips (dance and karate choreography) with exported RL policies.CMU Motion Capture Database — One of the oldest and most widely used motion-capture datasets in existence — 2,605 motion sequences across 6 categories from 140+Unitree UnifoLM-WBT — Unitree's whole-body teleoperation dataset for the G1 humanoid, released as a collection of task-specific Hugging Face repos ratheOmniRetarget — A human-motion-to-humanoid retargeting dataset for the Unitree G1 — about 4 hours of publicly released motion (object-interaction,
robotindex