Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement
COAP replaces runtime robot models with reusable state-tracking code, reporting 70.24% success on RoboDojo’s 42 bimanual tasks.
The paper frames robot control as an “Embodied Turing Machine,” where robot and environment state become the tape and policy code becomes the rule set. Its Code-Only-as-Policy library measures state from camera images and proprioception, then makes decisions without a VLM or VLA in the test-time loop. The authors argue this makes behavior explicit, cheaper to run, easier to recover, and extensible across tasks. They present it as a medium for recursive self-improvement because coding agents can modify the shared library in a closed loop. Source: HF Daily Papers' note
The paper frames robot control as an “Embodied Turing Machine,” where robot and environment state become the tape and policy code becomes the rule set. Its Code-Only-as-Policy library measures state from camera images and proprioception, then makes decisions without a VLM or VLA in the test-time loop. The authors argue this makes behavior explicit, cheaper to run, easier to recover, and extensible across tasks. They present it as a medium for recursive self-improvement because coding agents can modify the shared library in a closed loop. Source: HF Daily Papers' note
score 4