PartManip: Learning Cross-Category Generalizable Part Manipulation Policy from Point Cloud Observations

Haoran Geng, Ziming Li, Yiran Geng, Jiayi Chen, Hao Dong, He WangCVPR 2023 · OpenReview

TL;DR

This paper introduces PartManip, a new benchmark for training robots to manipulate objects across different categories using only point cloud data. They propose a novel two-stage approach: first training a state-based expert policy and then distilling its knowledge into a vision-based student policy, achieving significant improvements in cross-category generalization.

What Problem Does This Paper Address?

Building robots that can perform a wide range of manipulation tasks in real-world environments is a major challenge. Current approaches often struggle to generalize beyond the specific objects they were trained on. While previous work has focused on category-level manipulation (e.g., opening any drawer), this paper tackles the more ambitious goal of cross-category manipulation – teaching a robot to, for example, open a drawer on a storage unit and then generalize that skill to open a drawer on a microwave. This requires the robot to understand that similar parts (like handles or drawers) function similarly across vastly different object types, even when only observing partial, noisy sensor data like point clouds.

Methods

The authors propose a two-stage training pipeline:

  1. State-Based Expert Policy: They first train an "expert" policy that has access to "oracle" information, meaning it knows the precise state of the world, including the exact pose of object parts. To make learning easier, they introduce:
  2. Part-Based Canonicalization: This technique transforms all state information into a standardized coordinate frame relative to the target object part. This helps the policy focus on the relative relationships between the robot's gripper and the part, rather than absolute world coordinates, making it more robust to variations.
  3. Part-Aware Rewards: These rewards are designed to provide more informative feedback to the policy, guiding it towards successful manipulation by considering the state and movement of specific object parts.
  4. Vision-Based Student Policy (Knowledge Distillation): Once the expert is trained, its knowledge is transferred to a "student" policy that only uses realistic, partial point cloud observations. This is achieved through:
  5. DAgger (Dataset Aggregation): An online imitation learning algorithm that allows the student to learn from demonstrations collected by the expert, while also interacting with the environment and collecting new data.
  6. 3D Sparse UNet Backbone: A powerful neural network architecture designed to effectively process the complex and diverse point cloud data from various objects.
  7. Domain Adversarial Learning: This technique helps the student policy learn features that are invariant to object category, crucial for generalizing to unseen categories. It works by training a domain classifier to distinguish object categories and then using a gradient reversal layer to force the backbone to produce features that fool this classifier.
  8. Observation Augmentations: Techniques like point cloud jittering and color changes are applied during training to make the student policy more robust to variations in visual input.

Datasets

  • PartManip Benchmark: The paper introduces a new, large-scale benchmark called PartManip. It comprises 11 object categories, 494 objects, and 1432 tasks across 6 task classes (OpenDoor, OpenDrawer, CloseDoor, CloseDrawer, PressButton, GraspHandle). It's designed to be more diverse and realistic than existing benchmarks, using sparse-view point clouds as input without requiring oracle part segmentation.
  • GAPartNet: The object assets for PartManip are selected from the GAPartNet dataset, which provides rich annotations for generalizable and actionable parts (GAParts).

Key Insights and Ideas

  • Parts as the Key to Generalization: The core idea is that by focusing on manipulating "parts" (like handles, buttons, doors) that are common across different object categories, robots can learn more generalizable manipulation skills.
  • Two-Stage Learning is Crucial: Directly training a vision-based policy on complex, cross-category tasks with noisy sensor data is extremely difficult. A state-based expert provides a stable learning signal, and knowledge distillation effectively transfers this learned skill to a realistic vision-based agent.
  • Canonicalization Simplifies Learning: Transforming states into a part-centric coordinate frame significantly reduces the complexity and variability of the input, making it easier for the policy to learn.
  • Domain Adversarial Learning for Cross-Category Transfer: Explicitly training for domain-invariant features is essential for overcoming the significant visual differences between object categories.
  • Expressive Backbones Matter: The diversity of objects necessitates powerful neural network architectures like 3D Sparse UNet to extract meaningful features.

Key Results

  • Significant Performance Gains: The proposed PartManip benchmark and the two-stage learning method achieve state-of-the-art performance, outperforming baselines by a large margin, especially on unseen object categories. For example, in the OpenDoor and OpenDrawer tasks, their method shows a success rate improvement of over 20% on unseen categories compared to other methods.
  • Real-World Transfer: The method successfully transfers learned policies from simulation to a real-world robotic arm, demonstrating its practical applicability.
  • Ablation Studies Confirm Contributions: Experiments show that each component of their method (part-based canonicalization, domain adversarial learning, pre-training) contributes significantly to the overall performance, particularly for cross-category generalization.

Limitations and Future Work

  • Sim-to-Real Gap: The authors acknowledge the significant gap between simulation and real-world physics, which can still impact performance. They suggest using digital twin systems to mitigate this.
  • Further Improvement on Unseen Categories: While performance is strong, there's still room for improvement in handling entirely novel object instances and categories.
  • Benchmark for Future Research: The authors hope that the PartManip benchmark will serve as a valuable resource for future research in generalizable manipulation policy learning.

Summary updated

Abstract

Learning a generalizable object manipulation policy is vital for an embodied agent to work in complex real-world scenes. Parts, as the shared components in different object categories, have the potential to increase the generalization ability of the manipulation policy and achieve cross-category object manipulation. In this work, we build the first large-scale, part-based cross-category object manipulation benchmark, PartManip, which is composed of 11 object categories, 494 objects, and 1432 tasks in 6 task classes. Compared to previous work, our benchmark is also more diverse and realistic, i.e., having more objects and using sparse-view point cloud as input without oracle information like part segmentation. To tackle the difficulties of vision-based policy learning, we first train a statebased expert with our proposed part-based canonicalization and part-aware rewards, and then distill the knowledge to a vision-based student. We also find an expressive backbone is essential to overcome the large diversity of different objects. For cross-category generalization, we introduce domain adversarial learning for domain-invariant feature extraction. Extensive experiments in simulation show that our learned policy can outperform other methods by a large margin, especially on unseen object categories. We also demonstrate our method can successfully manipulate novel objects in the real world. Our benchmark has been released in https://pku-epic.github.io/PartManip.
BibTeX
@inproceedings{geng2023partmanip,
  author = {Haoran Geng and Ziming Li and Yiran Geng and Jiayi Chen and Hao Dong and He Wang},
  title = {{PartManip: Learning Cross-Category Generalizable Part Manipulation Policy from Point Cloud Observations}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2023},
  url = {https://openreview.net/forum?id=17LcJtWCtM},
}
Figure