Stanford researchers have introduced a robot controller that chooses its next move by matching the robot’s recent observations and actions to stored demonstrations. For three real-robot tasks, fitting the controller took 30 to 120 seconds on an NVIDIA RTX 4090; a fine-tuned π0.5 policy using the same demonstrations took 20 hours.
Called Behavior Predictive Control (BPC), the method keeps recorded examples available to the deployed controller instead of compressing them into a large neural policy. It searches for stretches of demonstrations with observations and actions similar to the robot’s recent history, combines the next actions from those stretches, then adjusts the result. “Training-free” means BPC does not train an end-to-end policy to choose actions; it still fits how it matches demonstration examples and adjusts their combined action.
On a UFactory xArm6 tabletop robot, BPC succeeded in 36 of 40 trials at moving a bag of beans onto a drawer and 38 of 40 trials on a sponge-in-drawer task. The π0.5 baseline scored 39 and 40. The clearest difference came on the “Switch” task, which required moving the bean bag to the plate opposite its starting side: BPC succeeded in 34 of 40 trials, compared with 8 of 40 for π0.5. The authors attribute the gap to history. Once the bag was in transit, π0.5 did not retain which side it came from; BPC used recent observations and actions to keep track.
In separate quadrotor tests, BPC ran onboard an 8 GB Jetson Orin Nano at 75 Hz and completed a figure-eight route in eight of 10 attempts. One failed run repeated part of the route; the other ended at the wrong gate. These results cover selected lab tasks. The authors say BPC depends on demonstrations covering the situation; the trials do not establish closed-loop stability or general safety, or show that BPC will complete tasks beyond the ones tested.
Sources
- Original paper: “Training-free Behavior Cloning”
- π0.5 baseline paper
- Lead image: Adang et al., Switch-task frame from the paper, CC BY-SA 4.0; image source.