Constrained Behavior Cloning for Robotic Learning

TL;DR

This work tackles the brittleness of behavior cloning in robotics due to limited perception and distribution shift by introducing GHCBC, a framework that fuses historically constrained BC (HCBC) with geometrically constrained BC (GCBC). HCBC encodes vision and action histories via a Transformer-based encoder with KL regularization to capture temporal structure, while GCBC imposes high-level geometric constraints on joint and end-effector poses to focus learning on relative pose information. The integrated Constrained Pose Transformer (GHCBC) predicts action sequences using vision, geometry, and history, employing action chunking and temporal ensembling, with history buffers that reset on gripper state changes. Empirical results on RLBench tasks with Sawyer robots show substantial improvements over a state-of-the-art BC baseline in both simulation and real-world settings, highlighting gains in robustness, stability, and long-horizon manipulation capabilities that are valuable for practical robotic imitation learning.

Abstract

Behavior cloning (BC) is a popular supervised imitation learning method in the societies of robotics, autonomous driving, etc., wherein complex skills can be learned by direct imitation from expert demonstrations. Despite its rapid development, it is still affected by limited field of view where accumulation of sensors and joint noise bring compounding errors. In this paper, we introduced geometrically and historically constrained behavior cloning (GHCBC) to dominantly consider high-level state information inspired by neuroscientists, wherein the geometrically constrained behavior cloning were used to geometrically constrain predicting poses, and the historically constrained behavior cloning were utilized to temporally constrain action sequences. The synergy between these two types of constrains enhanced the BC performance in terms of robustness and stability. Comprehensive experimental results showed that success rates were improved by 29.73% in simulation and 39.4% in real robot experiments in average, respectively, compared to state-of-the-art BC method, especially in long-term operational scenes, indicating great potential of using the GHCBC for robotic learning.

Figures (6)

• Figure 1: The framework of geometrically and historically constrained behavior cloning (GHCBC).
• Figure 2: Robotic arm pose and vector traction.
• Figure 3: History information with action chunking and temporal ensembling.
• Figure 4: Simulation tasks setup. (a) Sorting task setup. (b) Image observation in initial position of the sorting task. (c) Button pushing task setup. (d) Image observation in initial position of the button pushing task. (e) Basketball shooting task setup. (f) Image observation in initial position of the basketball shooting task. (g) Sorting task setup in real world.
• Figure 5: Simulation environment set up. (a) Wrist camera installation pose related to end-effector. (b) Multi-light setup in simulation scene.