Abstract
Behavior Cloning (BC) has emerged as a powerful paradigm for robot learning, yet it is often constrained by the availability of high-quality human demonstrations. While algorithmic experts like sampling-based motion planners can generate synthetic data at scale, standard methods such as Rapidly-exploring Random Trees (RRT) often produce high-entropy, inconsistent trajectories that are difficult for policies to imitate. This paper investigates the impact of data consistency on learning performance. We propose a "Planning-Guided Data Synthesis" pipeline that prioritizes low-entropy, predictable solutions. Our results demonstrate that policies trained on consistent demonstrations significantly outperform those trained on standard RRT data, achieving high success rates in complex bimanual and dexterous manipulation tasks with zero-shot transfer to real-world hardware.
Introduction
The success of Behavior Cloning in robotics is frequently bottlenecked by the difficulty of collecting human demonstrations for high-dimensional, contact-rich tasks. Coordinating 16 joints in a dexterous hand or synchronizing two robotic arms for large-object manipulation is non-trivial for human operators. Consequently, researchers have turned to algorithmic experts—such as Reinforcement Learning (RL) or sampling-based motion planning—to synthesize training data.
However, a critical challenge arises: Data Quality. Standard sampling-based planners like RRT are designed for probabilistic completeness, often exploring the state space randomly. This leads to high-entropy demonstrations where the same task is solved in numerous, meandering ways. For a student policy, this multi-modality is confusing and leads to poor convergence. We argue that for the purpose of teaching, consistency is more important than completeness.
Methods: Planning-Guided Data Synthesis
We propose a three-stage pipeline to generate high-quality, low-entropy demonstrations:
- Consistent Planning: We modify planners to favor predictable solutions. For bimanual tasks, we use a greedy search that minimizes regrasps. For dexterous manipulation, we utilize a Probabilistic Roadmap (PRM) with fixed canonical orientations and pre-computed primitives.
- Simulation Rollout: Plans are executed in a physics simulator to generate realistic state-action trajectories, capturing the dynamics of contact.
- Hindsight Goal Relabeling: To maximize data efficiency, we treat all trajectories—including partial successes—as demonstrations of reaching the final state achieved.
The student model is a Goal-Conditioned Diffusion Policy, chosen for its ability to model complex distributions while benefiting from the reduced noise of consistent demonstrations.
Experimental Results
Bimanual Manipulation
In the IiwaBimanual task, two KUKA Iiwa arms must rotate a large cylinder by 180°. This task requires precise coordination and frequent regrasping.
We compared a policy trained on "Greedy" (low-entropy) data against one trained on standard Contact-RRT data. The Greedy-trained policy achieved a 99% success rate with only 100 demonstrations. In contrast, the RRT-trained policy reached only 44% success with the same data volume and failed to exceed 84% even with 5,000 demonstrations. Furthermore, the Greedy policy transferred to real hardware with a 90% success rate, demonstrating robustness to physical discrepancies.
Dexterous In-Hand Manipulation
The AllegroHand task involves reorienting a cube to a target pose using a 16-DoF robotic hand. This environment is highly contact-rich and challenging for traditional BC.
Using a PRM-based global planner to ensure consistent movement primitives, we trained a policy that achieved a 62.5% success rate on real hardware via zero-shot transfer. The primary failure modes involved out-of-distribution states where the object fell into configurations unseen during training, suggesting a need for interactive imitation learning (e.g., DAgger) in future work.
Discussion and Conclusion
Our findings suggest a paradigm shift in how we generate synthetic data for robot learning. While traditional motion planning focuses on finding any path, generating data for imitation learning requires finding consistent paths. By reducing action entropy and making critical decisions—like regrasp timing—predictable, we can train highly effective policies for complex tasks that were previously difficult to master. Future work will explore extending these methods to dynamic manipulation and deformable objects.
Summary updated