Working Backwards: Learning to Place by Picking

Oliver Limoyo, Abhisek Konar, Trevor Ablett, Jonathan Kelly, Francois Robert Hogan, Gregory Dudek

TL;DR

Placing via Picking (PvP) is a self-supervised framework that allows robots to autonomously learn complex object placement tasks—like loading a dishwasher—without human demonstrations. By picking objects from their target locations and time-reversing the trajectories, the system generates high-quality training data for imitation learning, outperforming traditional human-guided (kinesthetic) teaching.

What Problem Does This Paper Address?

Imitation Learning (IL) is a powerful way to teach robots, but it has a massive bottleneck: it requires hundreds or thousands of "expert" demonstrations. For tasks like placing objects into tight, contact-constrained spots (e.g., a slot in a dish rack), collecting this data manually via kinesthetic teaching (physically moving the robot arm) is tedious, prone to human error, and doesn't scale.

While "picking" is a well-studied problem with robust planners, "placing" is much harder because it requires reasoning about environment constraints and precise alignments. This paper asks: Can we use the inherent symmetry between picking and placing to let the robot teach itself?

Methods

The authors introduce Placing via Picking (PvP), a four-stage autonomous data collection cycle:

  1. Grasp Planning: The robot uses a language-driven pipeline (Grounding-Dino and Segment Anything) to identify a specific object (e.g., "the green plate") and Contact-GraspNet to generate potential grasp poses while the object is already in its goal location.
  2. Grasping with Compliance (CCG): To ensure the object doesn't move before the "expert" trajectory is recorded, the robot uses Compliant Control for Grasping. This allows the arm to "give" slightly as it contacts the object, preventing it from being knocked out of its goal position.
  3. Tactile Regrasping (TR): Using Finger-STS visuotactile sensors, the robot checks if the grasp is stable. If the grasp is "shallow" or likely to slip, it autonomously adjusts its grip.
  4. Retrieving and Reversing: Once a stable grasp is achieved, the robot lifts the object and records the trajectory. By time-reversing this retrieval path, the robot creates a perfect "placing" demonstration.

The collected data is then used to train a vision-based policy (ResNet-18 backbone) via Behavioral Cloning. The authors also employ Noise Augmentation, intentionally perturbing the robot's pose during data collection to teach the policy how to recover from small errors.

Datasets

The study focuses on real-world experiments rather than standard simulated benchmarks:

  • Dishrack Loading: Placing plates of various sizes and materials into specific slots.
  • Table Setting: Placing a bowl onto a plate and a cup onto a coaster.
  • Comparison Dataset: A set of demonstrations collected via traditional kinesthetic teaching (a human moving the robot) to serve as a baseline.

Key Insights and Ideas

  • Symmetry as Supervision: The core "aha moment" is that a successful pick from a goal state is a perfect blueprint for a successful place into that state.
  • The Necessity of Compliance: Stiff robots are bad teachers. Without compliant control, the act of grasping an object in a tight slot often shifts the object or the rack, making the reversed trajectory inaccurate.
  • Programmatic > Human: Surprisingly, policies trained on PvP data outperformed those trained on human demonstrations. Human-provided trajectories were found to be longer, noisier, and less consistent than the robot's own reversed paths.
  • Tactile Filtering: Tactile sensing acts as an automated "quality control" for the dataset, ensuring that only demonstrations with stable, non-slip grasps are used for training.

Key Results

  • Autonomy: Using both CCG and TR, the system collected 128 consecutive demonstrations without a single human intervention. Without these modules, the system required human help every 9 episodes on average.
  • Performance: The PvP-trained policy achieved an 87.8% success rate on the challenging dishrack task using a Gaussian Mixture Model (GMM) action representation.
  • Efficiency: PvP policies consistently outperformed kinesthetic teaching policies by approximately 20% across various dataset sizes.
  • Robustness: Noise augmentation improved the final policy's success rate by about 10-15%, proving that "shaking" the robot during data collection helps it learn better error recovery.

Limitations and Future Work

  • Grasp Dependency: The method relies entirely on the ability of an off-the-shelf grasp planner to find a valid pick pose. If the object is in a position where it cannot be grasped (even if it could be placed), PvP cannot collect data.
  • Static Viewpoints: The camera viewpoint was manually selected to be "adequate." Future work could involve the robot autonomously moving the camera to find the best angle for data collection.
  • Tactile Integration: Currently, tactile data is used to collect the data, but it isn't an input for the final policy. Integrating tactile feedback into the real-time placing policy could further improve success in high-friction environments.

Summary updated

Abstract

We present placing via picking (PvP), a method to autonomously collect real-world demonstrations for a family of placing tasks in which objects must be manipulated to specific, contact-constrained locations. With PvP, we approach the collection of robotic object placement demonstrations by reversing the grasping process and exploiting the inherent symmetry of the pick and place problems. Specifically, we obtain placing demonstrations from a set of grasp sequences of objects initially located at their target placement locations. Our system can collect hundreds of demonstrations in contact-constrained environments without human intervention using two modules: compliant control for grasping and tactile regrasping. We train a policy directly from visual observations through behavioural cloning, using the autonomously-collected demonstrations. By doing so, the policy can generalize to object placement scenarios outside of the training environment without privileged information (e.g., placing a plate picked up from a table). We validate our approach in home robot scenarios that include dishwasher loading and table setting. Our approach yields robotic placing policies that outperform policies trained with kinesthetic teaching, both in terms of success rate and data efficiency, while requiring no human supervision.
BibTeX
@inproceedings{limoyo2024working,
  author = {Oliver Limoyo and Abhisek Konar and Trevor Ablett and Jonathan Kelly and Francois Robert Hogan and Gregory Dudek},
  title = {{Working Backwards: Learning to Place by Picking}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2024},
  doi = {10.1109/iros58592.2024.10802173},
}
Figure