ProgressCompass: context injection cuts embodied progress reward model error by 63%
Jianshu Zhang · hf · 2026-10-07
The paper tackles context-dependent progress estimation in long-horizon embodied tasks: Progress Reward Models (PRMs) score how far a task has come, but the current frame alone often can't tell, since progress depends on what happened before.
- Introduces ContextProgress-Bench: 24 manipulation tasks, 120 episodes, spanning State Recall, Sequence Tracking, and Recurrence Disambiguation settings
- Paired diagnosis shows even PRMs reading full history get lost, yet with the right context injected the same five models cut progress error by 77-82%
- Proposes ProgressCompass, an agentic loop using general-purpose VLMs to supply context to a frozen PRM, cutting error by 63% and raising rank agreement by 76%
Takeaway: embodied PRMs aren't incapable — they're lost without the right context.
More from Embodied
- Tesla's Cybercab rests on aggressive FMVSS interpretation NHTSA could reject — binarybits · 2026-10-07
- Waymo Colors Inside FMVSS Lines; Tesla Bets on Regulatory Favor, Analyst Argues — binarybits · 2026-10-07
- 19-Year-Old Founder Zain Raises $11M Seed Led by a16z to Ship Personal AI Computer Core — nick_linck · 2026-10-07
- Father-son team launches Rhem, a companion robot helping aging parents call, book and track health — ycombinator · 2026-10-07
- Sentdex finds decision language models fall short in real robotics pipelines vs RL/VLAs — Sentdex · 2026-10-07
- NeurIPS paper adds probabilistic uncertainty quantification to robot memory for better retrieval — lucacarlone1 · 2026-10-07