Robots Learn to Replan Tasks Based on Their Own Physical Strain

Most robot manipulation systems reason about the world outside the robot—object positions, obstacles, goals—while treating the robot's own physical condition as something only low-level controllers need to worry about. A new paper posted to arXiv (cs.RO) argues that's a mistake: a plan can look perfectly fine geometrically while quietly straining a joint or running into a mobility limit that makes it a poor choice in practice.

The researchers, led by Namiko Saito, introduce what they call body-grounded high-level replanning. In this framework, specific 'body-state events'—signals like rising joint load or restricted movement range—trigger a check-in at the strategy level, not just the control level. A large language model (LLM) is fed the robot's joint-level state, recent execution statistics, and a history of what's been tried so far. From this, it selects an alternative strategy suited to the robot's current physical condition, while the underlying task goal and low-level controller stay untouched.

As reported in the paper's abstract, the team tested this on a reaching task under two conditions: controlled external load and asymmetric mobility constraints, both in simulation and on a real robot. The body-grounded replanning approach maintained high task success rates while reducing the physical effort required, and it adapted strategies more efficiently than approaches without this feedback loop. Additional experiments on contact-rich manipulation tasks—beyond simple reaching—showed the same replanning interface could generalize to more complex interactions.

The broader takeaway, according to the authors, is that a robot's internal physical state shouldn't be confined to servo loops and safety limits. It can also feed into higher-level decisions about how a task gets done, not just whether it's geometrically possible. That distinction could matter as robots take on longer, more varied manipulation jobs where wear, fatigue-like load buildup, or partial mechanical constraints are common but rarely modeled explicitly in planning.