RobotUse: Allocating Computation, Context, and Decisions
RobotUse gives robot agents a structured way to revise physical actions without carrying every execution detail forward.
The harness has agents visually choose targets and poses while backend systems handle geometry, motion planning, and control. It uses subagents to keep detailed subgoal interactions separate, returning only what is needed for later decisions. The paper reports 45% task success on RoboLab, 6.7 points above CaP-X, with more compact decision context and less reliance on predefined action abstractions. The authors also say it can learn from imperfect real-world execution feedback and transfer that learning to later tasks. HF Daily Papers' note
The harness has agents visually choose targets and poses while backend systems handle geometry, motion planning, and control. It uses subagents to keep detailed subgoal interactions separate, returning only what is needed for later decisions. The paper reports 45% task success on RoboLab, 6.7 points above CaP-X, with more compact decision context and less reliance on predefined action abstractions. The authors also say it can learn from imperfect real-world execution feedback and transfer that learning to later tasks. HF Daily Papers' note
score 4