Teaching robots to work around their own broken joints

Most robots encounter physical degradation in real deployments—worn joints, weakened grippers, actuator failures—but existing vision-language-action models don't adapt to these constraints. This work adds a lightweight Health Projector module to VLA architectures that ingests a health vector encoding joint angles and torque limits, then adjusts predictions accordingly. Trained on 128 teleoperated episodes of degraded-robot behavior in the LIBERO environment, the model successfully completes spatial manipulation tasks with various joint configurations that the baseline pretrained model cannot handle. Code and dataset are to be released.