Robots that manipulate objects usually plan around the world — the positions of tools, obstacles and targets — while treating their own physical condition as a low-level control problem. A new study on arXiv, “Body-Grounded Replanning for Physically Adaptive Manipulation,” argues that a robot’s internal state (joint loads, mobility limits and recent execution history) can and should inform higher-level decisions about how to perform a task, not just how to control motors.

What the researchers propose

The paper introduces a framework the authors call body-grounded high-level replanning. Instead of only reacting at the motor-controller level when joints get overloaded or mobility becomes constrained, the system detects “body-state events” — signals that indicate the robot’s current physical condition is becoming problematic — and triggers a strategy-level replanner. A large language model (LLM) is used as the decision module: it interprets low-level sensor and joint information, recent execution statistics and the task execution history, and selects a context-appropriate alternative manipulation strategy. Crucially, the task objective (what the robot is trying to accomplish) and the underlying low-level controller remain unchanged; only the high-level plan is adapted.

Why this matters

Robots operating in real environments sometimes remain geometrically capable of a plan while becoming physically unsuitable to execute it — for example, a reaching motion that raises excessive joint torques or a trajectory that becomes inefficient due to asymmetric mobility after a partial actuator failure. By explicitly using internal body signals to inform strategy selection, the approach aims to preserve task success while reducing physical effort, limiting joint load and enabling more graceful adaptation to evolving physical constraints.