Fresh arXiv & Robotics Papers/Demos
Key Questions
What is Sanctuary GEN-1 robotics model?
Sanctuary GEN-1 achieves 99% dexterity at 3x speed using 500k+ hours of data for box folding and vacuum fixing. It handles disruptions and novel moves. This sets new reliability benchmarks.
What real-world application did DEEP Robotics deploy?
DEEP Robotics deployed LYNX M20 wheeled-legged robots for tea leaf transport in China's Zhejiang Province. This demonstrates practical quad robotics in agriculture. Deployment highlights commercialization.
What is Heracles in humanoid control?
Heracles bridges precise tracking and generative synthesis for general humanoid control. Released in Mar 2026, it advances diffusion-based methods. The paper details its approach.
What funding and hiring news in Chinese robotics?
Chinese firms secured funding: Spirit AI 1B yuan (3B total), UBTech offers $18M for chief AI scientist and released Walker S2. Moz JD focuses on coffee tasks. China pushes industrial standards for humanoid AI.
What is OpenAI's Safety Fellowship?
OpenAI launched a Safety Fellowship for independent AI research. It aims to foster safety-focused studies. This responds to growing embodied AI risks.
What is SKILL0 in RL?
SKILL0 enables in-context agentic RL for skill internalization. Released Apr 2026, it addresses scalable RL challenges. Video demo showcases its capabilities.
Why is data a bottleneck in robot control with LLMs?
Language models control robots effectively, but data scarcity limits progress, as per Avala analysis. π0 and others highlight this issue. Modular robots exacerbate the problem.
What safety advancements in HRI and robotics?
Advances include OpenClaw/GrabGen/ClawKeeper, collaborative HRI safety with LLMs, Harvard cy-trust, and US-led frameworks for robot networks. Physics-guided ML by Dr. Shao and CHAI/BADControl focus on safe embodied AI. Scenarios remain a bottleneck.
Sanctuary GEN-1 99% dex 3x speed 500k+ hrs data box folding/vacuum fix; DEEP Robotics tea transport quads; Heracles diffusion control; Stanford EgoNav G1; KAIST 7.3mph RL; X Square deploy; PhyAgentOS; OpenClaw/GrabGen/ClawKeeper safety; China funding (Spirit AI 1B yuan/3B total, Moz JD coffee); MMaDA-VLA 98%/Sharpa 73%/ThinkAct CALVIN; Unitree sports; UBTech $18M hire/Walker S2; OpenAI Safety Fellowship; Levine/Hwang talks; CHAI/BADControl; Harvard cy-trust; LLMs HRI; SkillX; scalable RL; Welch VLA; π0 data bottleneck; SKILL0 RL; China standards; GT-TD3 manip; MIT CSAIL multimodal; Dr. Shao physics-guided ML; collaborative HRI safety; scenarios bottleneck; Tokyo Summit May 2026.