BPC: A Training-Free Alternative to Neural Behavior Cloning

A new paper posted to arXiv proposes an alternative to the standard approach for teaching robots skills from demonstrations. Titled "Training-free Behavior Cloning," the work introduces a method called Behavior Predictive Control (BPC) that skips the usual step of training a large neural network on recorded demonstration data.

According to the paper's abstract, conventional neural behavior cloning compresses demonstrations into large models, which makes it hard to trace why the robot took a specific action and makes updating the policy expensive, since it typically requires retraining. Retrieval-based policies, which keep demonstrations accessible at runtime instead of baking them into network weights, avoid some of these issues but tend to struggle when live robot behavior diverges from what was recorded.

BPC aims to combine the strengths of both approaches. It uses three components: an action-aware retrieval metric to find relevant past demonstrations, a Hankel-based action-continuation prior — drawn from behavioral systems theory — to model how actions naturally continue over time, and a closed-form one-step residual correction to refine predictions. Rather than training end-to-end, BPC predicts the robot's next actions by blending segments of stored observation-action data that best reconstruct what has actually happened during the current run.

The authors, led by Maximilian Adang, report testing BPC across simulated benchmarks and on real robots. They state that BPC is competitive with learned policies such as π_0.5, a neural policy used as a baseline, and surpasses it in some cases. The efficiency gains are notable: policy fitting drops from hours to seconds on consumer GPUs, and the system supports closed-loop control at up to 75 Hz on a Jetson Orin Nano, an embedded computing board commonly used for on-robot inference.

Beyond raw performance, the paper highlights two side benefits of keeping demonstrations in the loop. First, the retrieved demonstration windows and the coefficients used to blend them offer a built-in estimate of how far along the robot is in completing a task. Second, because predictions remain traceable to the specific demonstrations that produced them, engineers can inspect and even revise robot behavior simply by editing the demonstration bank, rather than retraining a model.

As reported in the arXiv listing, the paper was submitted on September 24, 2026, and is filed under the Robotics (cs.RO) category. It has not yet undergone formal peer review, and full experimental details, including exact benchmark results and hardware setups, are only available in the full PDF rather than the abstract summarized here.

If the reported results hold up under scrutiny, BPC's central pitch — competitive performance with far cheaper updates and built-in interpretability — could appeal to robotics teams that need to iterate quickly on new skills without the overhead of full model retraining.