Planning-Guided Diffusion Policy Learning for Generalizable Contact-Rich Bimanual Manipulation (2412.02676v2)

Abstract: Contact-rich bimanual manipulation involves precise coordination of two arms to change object states through strategically selected contacts and motions. Due to the inherent complexity of these tasks, acquiring sufficient demonstration data and training policies that generalize to unseen scenarios remain a largely unresolved challenge. Building on recent advances in planning through contacts, we introduce Generalizable Planning-Guided Diffusion Policy Learning (GLIDE), an approach that effectively learns to solve contact-rich bimanual manipulation tasks by leveraging model-based motion planners to generate demonstration data in high-fidelity physics simulation. Through efficient planning in randomized environments, our approach generates large-scale and high-quality synthetic motion trajectories for tasks involving diverse objects and transformations. We then train a task-conditioned diffusion policy via behavior cloning using these demonstrations. To tackle the sim-to-real gap, we propose a set of essential design options in feature extraction, task representation, action prediction, and data augmentation that enable learning robust prediction of smooth action sequences and generalization to unseen scenarios. Through experiments in both simulation and the real world, we demonstrate that our approach can enable a bimanual robotic system to effectively manipulate objects of diverse geometries, dimensions, and physical properties. Website: https://glide-manip.github.io/

• The paper introduces G-PLIDLE, a novel planning-guided diffusion policy learning method for generalizable contact-rich bimanual manipulation.
• It synthesizes high-quality training data using a model-based planner with smoothed contact models and filters for efficient data generation.
• The method employs a diffusion policy framework with task-conditioned learning, residual actions, and augmentation for robustness and generalization to out-of-distribution scenarios.

Planning-Guided Diffusion Policy Learning for Generalizable Contact-Rich Bimanual Manipulation

This paper introduces a novel method, \MethodAcronym~(Generalizable PLanning-GuIded Diffusion Policy LEarning), which focuses on solving the challenges associated with contact-rich bimanual manipulation tasks. Such tasks are inherently complex due to the need for precise coordination between two robotic arms to manipulate objects through dynamic multi-contact interactions. The authors address two primary challenges: generating sufficient high-quality demonstration data and effectively generalizing learned policies to unobserved scenarios.

Core Contributions

1. Data Synthesis via Model-Based Planning: The paper utilizes a model-based motion planner, leveraging privileged information available in high-fidelity physics simulations, to synthesize data efficiently. This approach capitalizes on recent advancements in smoothed contact models to generate trajectories while maintaining computational efficiency. The method involves filtering, wherein only high-quality trajectories are retained for training, ensuring that the simulation-to-reality gap is minimized.
2. Diffusion Policy Learning: Building on the synthesized data, the paper adopts a diffusion policy framework, which excels at learning from complex, multimodal data. The authors introduce several innovations to enhance the robustness and generalization of policies. Notably, they utilize task-conditioned learning and residual joint actions to predict sequences of actions tailored to the specific task requirements, along with data augmentation strategies like the Flying Point Augmentation to tackle real-world noise and discrepancies.
3. Feature Extraction and Task Representation: The policy architecture incorporates point cloud processing to capture essential geometric features from visual observations. The task's dynamic nature is further encapsulated by representing tasks as transformations relative to initial object states, which aids in focusing the learning process on goal-oriented manipulations without presupposing known object geometries.

Experimental Evaluation and Results

The paper presents rigorous evaluations of \MethodAcronym, demonstrating its efficacy both in simulated environments and real-world settings. The policy's performance is benchmarked against different rotational tasks, demonstrating significant improvement over baseline performance. The evaluations reveal:

• In-Distribution Success: The policy trained via \MethodAcronym excels in in-distribution tasks, achieving notable success rates and outperforming traditional planners, especially in scenarios requiring intricate multi-phase contacts.
• OOD Generalization: Impressively, the learned policy retains robustness and adaptability when tested on out-of-distribution geometries and material properties, including soft, deformable, and irregularly shaped objects. These evaluations underscore the approach's potential for real-world applicability beyond the conditions encapsulated in the training dataset.
• Adaptiveness to Novel Environments: Through the integration of visual point-cloud observations and real-time task adaptation, \MethodAcronym demonstrates strong potential for real-world deployment by effectively closing the gap between simulation and reality.

Implications and Future Directions

The implications of this research are both practical and theoretical. Practically, the method sets the stage for extending robotic manipulation into more generalized tasks, potentially impacting multifarious sectors from logistics to personal robotics. Theoretically, it offers insights into effective learning strategies for multimodal policies in robotics, showing the synergistic benefits of combining classical planning with modern data-driven approaches.

Future work, as suggested by the authors, might consider scaling this approach to even more complex dexterous tasks, incorporating richer datasets with diverse object characteristics. Additionally, the integration of this method with real-time adaptive planning could further enhance its utility in dynamic environments, potentially paving the way for breakthroughs in autonomous robotic systems.

Conclusion

In conclusion, this paper underscores the strength of a hybrid approach that combines planning and learning to achieve generalizable bimanual manipulation. \MethodAcronym represents a promising advance in the field, successfully bridging simulation and real-world application gaps while showcasing the potential of diffusion policies in contact-rich manipulation tasks. It presents a significant contribution to robotic manipulation, proposing a scalable path forward for complex real-world interactions.