Mind2ActWorld: Tsinghua-led benchmark dissects why robots fail, step by step
机器之心 · wechat · 2026-11-22
Amid the 'VLA/WAM dead, foundation-model agents rising' debate, Tsinghua mmlab@SIGS and Velabot AI plus 16 institutions released Mind2ActWorld: 30 tasks, 180 configs, testing 'mind' (memory/reasoning), 'act' (physical execution), and their composition separately, with per-episode failure attribution and diagnosis cards. Across 9 systems and 56,700 sim episodes, the best agent (GPT-6 Astra) scores 52.7% on Mind, 39.0% on Act, only 12.0% on Mind2Act; end-to-end policies top out at 3.6%. Most failures break in physical execution, not coordination. The report also derives data/tooling fixes from breakpoints and counts hidden costs: 11-35.6s per agent call, $25.13 per episode API cost. Verdict: task-specific models plus self-evolution loops, not one winner.
More from Embodied
- LynnReal-Hand: hand-controlled video world model streams 540p at 18 FPS on one H200 — _akhaliq · 2026-11-23
- 25 Ways to Make VLA Robots Faster, Mapped From Model to Server — TheTuringPost · 2026-11-23
- Arrowmax's AI screwdriver S1 auto-detects screws and torque, from $129 — Scobleizer · 2026-11-23
- Figure 04 mass production next year; Tesla Optimus V3 polished through 2027 — ChrisGPT · 2026-11-23
- Musk touts Tesla superintelligence, chip design & Terafab as shared world model powers Autopilot and Optimus — elonmusk · 2026-11-23
- Ashlee Vance's decade-long Neuralink reporting becomes LIFE magazine's relaunch cover story — kyliebytes · 2026-11-23