New Frontiers in Humanoid Robotics and the Physical AI Ecosystem

로보틱스 | Wed Jun 17 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 6 sources

From Genesis AI's non-humanoid humanoid robot to robot training data infrastructure and LeRobot integration SDKs, physical AI is expanding across multiple fronts.

Analysis

[Genesis AI] unveiled general-purpose humanoid robot Eno ^[1]

• Non-humanoid design without head or legs
• Foldable structure on a wheeled base
• Equipped with hands identical in form and function to human hands
• Targeting deployment in manufacturing
• labs
• and logistics from late 2026

[XDOF] exited stealth and raised $70 million for robot training data infrastructure ^[2]

• Investors include Thrive Capital
• Spark Capital
• a16z
• Lux
• and WndrCo
• Working with 20 customers including frontier AI labs
• Data collection based on GELLO teleoperation system
• Plans to release the largest robot learning dataset with UC Berkeley AI Research lab

[IO-AI Tech] operated a VR-based humanoid robot teleoperation and training data collection system ^[6]

• Chinese startup located near Shenzhen
• Remote control of humanoids via VR headsets and motion tracking equipment
• Simultaneous control of robot hands from 10 manufacturers with the same glove
• Partnerships with convenience store chains and Jack Sewing Machines

[AWS Strands Robots] released an open-source SDK integrating Hugging Face Hub and LeRobot ^[4]

• Open-sourced under Apache 2.0 license
• Exposes the LeRobot stack as AgentTool
• Simulation and real hardware use the same data format
• Coordinates multi-robot fleets via Zenoh peer mesh
• Supports policy inference for GR00T and MolmoAct2

[Reachy Mini] added MCP-based remote tool calling functionality ^[5]

• Calls tools hosted on Hugging Face Spaces via MCP
• Tools can be added with a single command
• Runs in the Space without downloading code
• Active tools controlled via tools.txt in profiles

[OpenRouter] compared model behavior characteristics through an 11-LLM battle royale experiment ^[3]

• Grok 4.1 Fast won with 13 victories out of 30 games
• Cost per win of $0.97
• 27x more efficient than Claude Sonnet 4.6
• Claude Sonnet 4.6 showed team-forming and cooperative tendencies
• GPT 5.4 had the most kills but only 2 wins
• Existing benchmarks failed to predict wins and losses

Sources

• [1] The next humanoid robot might not look human at all - The Verge AI
• [2] Collecting robot training data is dirty, unglamorous work. Some AI labs are already paying XDOF to do it. - TechCrunch AI
• [3] A robot is sprinting towards you. Do you want it running on Claude or Grok? - Hacker News
• [4] From the Hugging Face Hub to robot hardware with Strands Agents and LeRobot - Hugging Face Blog
• [5] Adding MCP Tools to Reachy Mini - Hugging Face Blog
• [6] Operating a Humanoid With Your Body Is a Hot Job in China’s Hardware Capital - Wired AI