Explore the Ecosystem

Chinese robotics company developing humanoid, wheeled-legged, and quadruped robots for industrial and specialized applications.

Consumer electronics and AI company developing embodied AI and robotics foundation models.

Leading Chinese embodied AI and robotics company founded in 2023, developer of full-size humanoid robots, industrial manipulation systems, and dexterous hands.

Boston Dynamics is an American engineering and robotics design company known for highly dynamic mobile robots including Atlas, Spot, and BigDog. Owned by Hyundai Motor Group.

Swiss robotics company (ETH Zurich spin-off) developing full-stack dexterous manipulation — robotic hands, wearable data collection devices, and physical AI models for industrial automation.

Robotics research lab under Tencent, focused on embodied AI foundation models and real-world robot systems. The team behind the HY-Embodied family and HyVLA robot learning stack.

UBTECH Robotics (优必选科技) is a Chinese AI and robotics company specializing in humanoid robot development for industrial and consumer applications, listed on the Hong Kong Stock Exchange.

Google DeepMind is the AI research division of Google, formerly DeepMind Technologies, focused on fundamental AI research and embodied intelligence through robotics foundation models.

NVIDIA is the world's leading AI computing company, providing GPU-accelerated hardware and software platforms for AI, robotics, and simulation, including the Isaac robotics platform and Jetson edge AI hardware.

Unitree Robotics is a Chinese robotics company and the world's largest producer of humanoid robots by volume, known for the G1 and H1 humanoids at disruptive price points.

Physical AI / Embodied AI foundation model company, known for developing the Xingdong humanoid robot series.

General-purpose embodied AI company developing cross-embodiment foundation models for robotic manipulation, founded in 2024 by former Apple AI Platform lead Tian Ye.

Transformable humanoid and quadruped robot from Unitree Robotics, capable of carrying a human operator.

Heavy-payload force-controlled industrial robot from AGIBOT for material handling and palletizing.

Modular humanoid development platform from AGIBOT for education, research, and competitions.

Full-size humanoid robot from AGIBOT designed for commercial and service applications, with 51 DOF and NVIDIA Thor compute.

Flagship bipedal humanoid robot platform from Booster Robotics, designed for embodied AI development and research with strong onboard compute and open ecosystem.

A tendon-driven anthropomorphic robotic hand with bidirectional pulley-guided tendons and tactile fingertip sensors, designed for industrial automation and AI-first learning.

UBTECH full-size mass-produced ultra-bionic humanoid companion robot with 88 DOF, emotion-aware LLM, Agent Memory OS, and fast-slow brain dual architecture.

Franka FR3 is a research-oriented robotic arm from Franka Robotics with torque control, force interaction, and 7 degrees of freedom. Widely used in academic manipulation research and embodied AI benchmarks.

UR10e is a collaborative robot arm from Universal Robots with 10 kg payload capacity, 1300 mm reach, and 6 degrees of freedom. Widely used in robotics research, industrial automation, and manipulation benchmarks.

Unitree A2 is an industrial-grade quadruped robot with 100 kg standing load, 180 N·m joint torque, dual battery hot-swap, and IP56 rating for logistics, inspection, and rescue. Available as A2 (1 LiDAR) and A2-PRO (2 LiDAR, IP67).

Unitree Go1 is a consumer-grade quadruped robot dog launched in 2021, featuring Intelligent Side-Follow System (ISS), Super Sensory System (SSS), and a top speed of 4.7 m/s.

Wheel-legged quadruped robot developed by Unitree Robotics, combining wheeled mobility efficiency with legged obstacle-crossing capability.

38-billion parameter multimodal autoregressive foundation model from Xiaomi for embodied AI, unifying scene generation, trajectory transfer, and video generation.

VLA foundation model from Xiaomi for robot manipulation, trained on over 100K hours of real-world trajectory data.

An expressive and scalable transformer architecture for whole-body humanoid control, reducing MPKPE by 82% in global mode, validated in real-world deployment.

A real-time streaming inference framework for flow-matching VLA policies, exploiting timestep-invariance for 2.58x speedup and 50Hz stable streaming. Accepted at ICML 2026.

An action-conditioned world model from MIT and Harvard using drifting generative models to generate future frames at 30+ fps, 17x faster than diffusion baselines.

Test-Time-Training Robot Policies — a VLA foundation model from NVIDIA Research, Stanford, and University of Texas at Austin that scales visuomotor context to 8,000 timesteps with constant inference latency.

A 4-billion parameter compact world model from NVIDIA for on-device vision reasoning and robot control, built on the Nemotron architecture.

GigaWorld-Policy-0.5 is an action-centered World Action Model with Mixture-of-Transformers architecture, achieving 85ms inference latency on RTX 4090 for robot control.

End-to-end Vision-Language-Action (VLA) policy built on the HY-Embodied-0.5 backbone, spanning data collection, continued pre-training, SFT, RL post-training, and real-world deployment.

A family of embodied foundation models (VLM) with MoT architecture, 2B and 32B parameter variants, designed for spatial and temporal visual perception and embodied reasoning.

LingBot-VA is a World-Action Model (WAM) using autoregressive diffusion for simultaneous visual dynamics prediction and action inference in robot control.

π0.5 (pi-zero point five) is Physical Intelligence's upgraded VLA foundation model with improved open-world generalization trained via knowledge insulation on 10,000+ hours of robot data.

A benchmark dataset for capturing hand grasping of objects, providing 2D/3D annotations for hand pose estimation, 6D object pose estimation, and robot-to-human object handover tasks.

A richly-annotated, large-scale repository of 3D CAD models organized under the WordNet taxonomy, containing millions of models across thousands of semantic categories with consistent rigid alignments, parts, and symmetry annotations.

A large dataset of high-fidelity 3D objects with 800K+ annotated 3D models across diverse categories, designed to support 3D vision, embodied AI, and simulation research.

A large-scale egocentric video benchmark recorded by 32 participants in their native kitchen environments, capturing 100 hours of non-scripted daily kitchen activities with 90K+ action segments.

A massive-scale egocentric video dataset with 3,670 hours of daily-life activity video spanning hundreds of scenarios, captured by 931 unique camera wearers from 74 worldwide locations across 9 countries.

A framework for procedural generation of Embodied AI environments, enabling sampling of arbitrarily large datasets of diverse, interactive, customizable virtual environments for navigation, interaction, and manipulation tasks.

An artist-authored, annotated, reconfigurable 3D dataset of apartments with articulated objects (cabinets, drawers, etc.) designed for training virtual robots in interactive physics-enabled environments, part of Habitat 2.0.

A dataset of 18 highly photo-realistic 3D indoor scene reconstructions with dense meshes, HDR textures, semantic class and instance information, and planar mirror/glass reflectors for embodied AI research.

An RGB-D video dataset containing 2.5 million views across 1,513 indoor scenes annotated with 3D camera poses, surface reconstructions, and semantic segmentations for 3D scene understanding research.

A large-scale RGB-D dataset containing 10,800 panoramic views from 194,400 images of 90 building-scale indoor scenes with surface reconstructions, camera poses, and 2D/3D semantic annotations.

BridgeData V2 is a large-scale robot manipulation dataset from UC Berkeley RAIL Lab with 60,000+ WidowX trajectories across diverse scenes for training generalist robot policies.

Largest high-resolution dataset of real-world 3D indoor environments with 1,000 spaces covering 215 million m², designed as the backbone for embodied AI navigation and interaction research.

REAL is an agentic framework for open-world mobile manipulation, accepted to ECCV 2026, with REAL-Bench (241 tasks) and 78.3% end-to-end success on a physical dual-arm mobile robot.

GigaWorld is an umbrella project developing World Action Models (WAMs) for robot policy learning, including GigaWorld-0, GigaWorld-Policy, GigaWorld-Policy-0.5, and GigaWorld-1.

A family of embodied foundation models for real-world agents, developed by Tencent Robotics X and HY Vision Team. Encompasses embodied VLMs (HY-Embodied-0.5) and VLA policies (HyVLA-0.5).

NVIDIA Isaac GR00T is an open reference platform and development project for general-purpose humanoid robotics, combining open foundation models (GR00T N1 series), data pipelines, simulation frameworks, CUDA-X accelerated runtime libraries, and NVIDIA Jetson Thor for real-time robot inference and control.

A serving system for Robot Foundation Models (RFMs) designed for robot factories, providing GPU-pool serving, robotics-aware programming abstractions, and factory-oriented scheduling.

A portable inference runtime for deploying embodied AI foundation models across heterogeneous hardware platforms.

Agentic Skill Programming through Iterative Robot Exploration (ASPIRE) is a continual robot learning framework that enables robots to iteratively execute tasks, diagnose failures, repair robot control programs using large language models, and accumulate reusable skills in a continually expanding skill library.

An explicit 3D scene representation based on anisotropic 3D Gaussians for high-quality real-time radiance field rendering and novel view synthesis.

Training-free neurosymbolic manipulation framework using sequential planning via anchored robotic keypoints, achieving 43.7% on six LIBERO-PRO position and task evaluation cells.

Vision-language-kinematics synthetic data generation framework for perception-based humanoid loco-manipulation from reconstructed scenes.

GPU-accelerated photorealistic simulation platform from Stanford for household-level embodied AI tasks, powering the BEHAVIOR-1K benchmark.

Pioneering photorealistic simulation environment for embodied AI with 1,400+ real-world 3D indoor spaces, enabling virtual-to-real transfer for navigation and interaction.

LIBERO (Lifelong Robot Learning Benchmark) is a benchmark suite for studying knowledge transfer in lifelong robot manipulation, featuring 130 tasks across five task suites with standardized evaluation protocols.

A unified sim-and-real benchmark for evaluating generalist robot manipulation policies, with 42 simulation tasks and 18 real-world tasks. Built on Isaac Sim.

BEHAVIOR-1K is a comprehensive embodied AI benchmark with 1,000 everyday activities across 50 scenes and 9,000+ objects, built on the OmniGibson simulation environment for human-centered robotics research.

LIBERO-Pro is a robotic manipulation benchmark designed to evaluate long-horizon task generalization and skill composition.