MicroPlanBench: no SOTA multimodal model tops 70% on next-subtask judgment for robots
animesh_garg · x · 2026-10-21
Heimmoritz's team introduces MicroPlanBench, testing a key skill for robot brains: judging which subtask to do next in long-horizon goals.
Key findings:
- No SOTA multimodal model scored above 70%, meaning goals spanning 13+ subtasks would succeed less than 1% of the time
- Surprisingly, providing the task goal did not improve next-task prediction
- The team sees this as a critical gap for physical AI models and will release more benchmarks correlated with real-world success, with plans to open-source
Related event: MicroPlanBench: No SOTA VLM Beats 70% on Robot Next-Step Planning(3 posts)→
More from Embodied
- OpenTac auto-routes tactile sensor PCBs from a 3D model in 3 seconds, covering 93% of surface — anand_bhattad · 2026-10-23
- fal starts fine-tuning a physical robot to draw, results next week — OdinLovis · 2026-10-23
- PEARL demo: an LLM agent that checks experiment results, revises plans, and re-tests — KevinKaichuang · 2026-10-23
- CU Boulder's ATLAS Institute Opens Tenure-Track Robotics Professorship — ryosuzk · 2026-10-23
- MiTaS multi-resolution tactile imitation learning hits 80% success vs 35.5% vision-only baseline — Jan_R_Peters · 2026-10-23
- Princeton's Action Self-Distillation Teaches Robots Language Navigation with Zero Human Labels — shahdhruv_ · 2026-10-23