Physical AI & Embodied Intelligence¶
AGI cannot exist only in text -- it must understand and act in the physical world. Physical AI bridges foundation models with embodiment: humanoid robots, manipulation, navigation, and physics simulation. Moravec's Paradox -- sensorimotor skills are harder than abstract reasoning -- remains one of the deepest unsolved AGI challenges.
Humanoid Robotics & AGI Hardware¶
The race to build general-purpose humanoid robots capable of operating in unstructured human environments. These platforms are the physical embodiment layer for AGI.
| Company | Description | Links |
|---|---|---|
| Figure AI | Humanoid robots with advanced AI. Partnered with OpenAI and Microsoft for embodied AI. Figure 02 demonstrates autonomous manipulation, language-guided task execution, and learning from human feedback. Raised $2.6B+. | figure.ai |
| Tesla Optimus | General-purpose humanoid robot leveraging Tesla's massive autonomous driving AI stack (FSD neural nets, Dojo supercomputer). Targeting manufacturing and consumer deployment at scale. | tesla.com |
| Boston Dynamics | Pioneers of advanced locomotion. Atlas (humanoid) and Spot (quadruped) set the standard for physical capability, agility, and dexterity. Now Hyundai-owned, pivoting to AI-first control. | bostondynamics.com |
| 1X Technologies | NEO humanoid robot with human-like form factor. Backed by OpenAI. Focused on safe, embodied AI for real-world deployment in homes and workplaces. | 1x.tech |
| Unitree Robotics | Democratizing humanoid robotics with affordable platforms (G1, H1 series). Enables broad research access to embodied AI experimentation. | unitree.com |
| Sanctuary AI | Phoenix humanoid robot powered by "Carbon" -- a proprietary AI system designed for general-purpose task learning and autonomous execution. | sanctuary.ai |
| Agility Robotics | Digit bipedal robot designed for logistics and warehouse automation. Real-world deployment partner with Amazon. | agilityrobotics.com |
| Apptronik | Apollo full-size humanoid robot for industrial applications. Emphasis on safe human-robot collaboration. Mercedes-Benz partnership. | apptronik.com |
Robot Foundation Models (Vision-Language-Action)¶
The frontier of embodied AI research: models that take visual observations and language commands as input and directly output robot motor actions. VLA models represent the convergence of foundation models and physical intelligence.
| Model | Org | Year | Description | Links |
|---|---|---|---|---|
| Gemini Robotics 1.5 | Google DeepMind | 2025 | Agentic VLA model that turns visual information and language instructions into motor commands. Generality across novel situations, dexterity (origami, food prep), agentic tool-use, and thinking before acting. Supports multiple embodiments (ALOHA, Franka, Apptronik Apollo). Dual approach with Robotics-ER 1.5 for embodied reasoning. | Site, Report |
| pi0 | Physical Intelligence | 2024 | VLA flow model for general robot control. Novel flow matching architecture on top of pre-trained VLM. Trained on diverse dexterous tasks (laundry folding, table cleaning, box assembly) across single-arm, dual-arm, and mobile manipulators. Open-sourced weights. | Paper, Site |
| RT-2 | Google DeepMind | 2023 | Vision-Language-Action model that transfers web-scale knowledge to robotic control. Expresses actions as text tokens, enabling emergent reasoning (pick up the "improvised hammer" -> picks rock). 6k evaluation trials. | Paper |
| PaLM-E | 2023 | 562B-parameter embodied multimodal language model. Directly incorporates continuous sensor modalities into LLMs. Positive transfer across internet-scale language, vision, and robotics domains. | Paper | |
| Open X-Embodiment / RT-X | 21 Institutions | 2023 | Largest robotics dataset: 22 robots, 527 skills, 160k+ tasks from 21 institutions. RT-X model shows positive transfer across robot morphologies. The "ImageNet moment" for robotics. | Paper |
| OpenVLA | Stanford / UC Berkeley | 2024 | Open-source 7B-parameter VLA. Democratizes embodied AI research -- matches proprietary models on manipulation benchmarks. Fine-tunable for new robots and tasks. | GitHub |
| NVIDIA GR00T | NVIDIA | 2024 | Foundation model for humanoid robots. Multimodal inputs (text, video, demonstration) to robot actions. Part of NVIDIA's Physical AI platform alongside Isaac and Cosmos. | nvidia.com |
Simulation & Infrastructure for Physical AI¶
Training embodied AI requires massive simulation before real-world deployment. These platforms enable sim-to-real transfer, digital twins, and scalable robot learning.
| Platform | Description | Links |
|---|---|---|
| NVIDIA Isaac Sim / Isaac Lab | Production-grade robotics simulation platform with photorealistic rendering, physics accuracy, and domain randomization. Isaac Lab provides GPU-accelerated RL environments for robot learning at scale. | Developer, GitHub |
| NVIDIA Omniverse | Collaborative 3D simulation platform for building digital twins and physics-based robotics simulation. Foundation for NVIDIA's Physical AI ecosystem. | nvidia.com/omniverse |
| MuJoCo | Google DeepMind's open-source physics engine optimized for robotics and biomechanics. Fast, accurate contact dynamics. The standard tool for embodied AI research and RL benchmarking. | mujoco.org, GitHub |
| Genesis | Next-generation open-source physics engine for embodied AI. Differentiable simulation enabling gradient-based learning for physical systems. | GitHub |