Training-free Behavior Cloning

Training-free Behavior Cloning introduces Behavior Predictive Control, or BPC, a robot imitation method that avoids end-to-end policy training by retaining raw demonstration trajectories and combining three components at runtime: action-aware retrieval, regularized Hankel-based history reconstruction, and a closed-form residual correction based on Random Fourier Features. Unlike Diffusion Policy, OpenVLA, or the vision-language-action model π0.5, BPC keeps every prediction traceable to supporting demonstrations, allowing operators to edit the demonstration bank, estimate task progress, and deploy across different robots and modalities, including RGB, proprioception, 3D inputs, and raw pixels. On MimicGen tabletop simulations using 1000 training demonstrations, BPC achieved a 73 percent mean success rate, competitive with π0.5 at 77 percent, while fitting in 1–5 minutes rather than hours. On real xArm6 manipulation, fitting took 30–120 seconds versus 20 hours for π0.5; notably, BPC solved the memory-dependent Switch task in 34 successes out of 40, compared with 8 out of 40 for π0.5. The method also controlled a drone onboard a Jetson Orin Nano at 75 Hz, demonstrating deployment on resource-constrained hardware. Its limitations are demonstration coverage, storage and retrieval costs, and vulnerability to ambiguous or out-of-distribution histories, but the central result is a modular behavior-cloning policy that links inference directly to editable data instead of opaque neural weights.