A System for Imitation Learning of Contact-Rich Bimanual Manipulation Policies

Simon Stepputtis, Maryam Bandari, Stefan Schaal, Heni Ben Amor

TL;DR

This paper introduces a robust framework for teaching robots complex, bimanual (two-armed) tasks through human demonstration. By combining Bayesian Interaction Primitives with admittance control, the system allows robots to sense physical resistance and adapt their timing and motion in real-time, achieving a 90% success rate in a challenging multi-point assembly task.

What Problem Does This Paper Address?

Most robot imitation learning focuses on single-arm movements in "clean" environments where the robot doesn't touch much. However, real-world assembly—like inserting a bracket onto four tight-fitting pins—is "contact-rich."

In these scenarios, two major hurdles arise:

  1. Coordination & Internal Forces: Two arms holding the same object can easily fight each other, causing oscillations or hardware damage.
  2. Temporal Uncertainty: If a robot hits a snag or a physical obstruction, a standard "play-back" of a recorded motion will fail because the robot's internal clock keeps ticking while its physical progress has stopped.

The authors aim to create a system that doesn't just mimic where a human moved, but understands the relationship between force, space, and time.

Methods

The system uses a "stacked" control architecture:

  • Ensemble Bayesian Interaction Primitives (EnBIP): This is the "brain." It treats a task not just as a sequence of positions, but as a joint probability distribution of spatial data (positions) and temporal data (phase). It uses an Ensemble Kalman Filter (EnKF) to constantly update its belief about where it is in the task based on current sensor readings.
  • Dynamic Phase Estimation: Instead of following a fixed timer, the robot estimates its "phase" (0% to 100% completion). If the robot feels a counter-force (sensed via Force/Torque sensors), the phase estimation can slow down, stop, or even move backward until the obstruction is cleared.
  • Admittance Control: This is the "muscle" layer. It makes the robot compliant. Instead of rigidly following a path, the robot acts like a spring-mass-damper system, allowing it to "give" slightly when it encounters unexpected resistance, preventing damage and allowing for smoother insertions.

Datasets

The researchers evaluated their system using:

  • Simulated Environment: A MuJoCo simulation of a UR5 robot performing a two-pin insertion with 1mm and 5mm tolerances.
  • Real-World Bimanual Setup: A dual-robot station (UR5 and UR10) tasked with inserting a bracket onto four pins.
  • Human-Subject Study: 8 participants provided 30 demonstrations each using two different methods: Kinesthetic teaching (physically moving the robot arms) and Space-Mouse teleoperation (using a 6-DoF joystick).

Key Insights and Ideas

  • The "Phase" is Paramount: The most critical finding is that spatial accuracy isn't enough. For contact-rich tasks, the robot must be able to "localize in time." If the bracket gets stuck on a pin, the robot needs to realize it hasn't progressed and adjust its motion accordingly rather than blindly pushing forward.
  • The Teleoperation Trade-off: The study revealed a fascinating dichotomy in data collection. Kinesthetic teaching is much easier for humans (lower mental load) and produces "smoother" looking paths. However, Space-Mouse teleoperation—despite being twice as labor-intensive for the human—provides "cleaner" Force/Torque data. In kinesthetic teaching, the sensors pick up the human's hands pushing the robot; in teleoperation, the sensors only pick up the interaction between the robot and the environment.
  • Force Data is the "Secret Sauce": While the robot can learn the basic motion from joint angles alone, success rates in varied environments jump significantly (from ~70% to 90%) when Force/Torque data is included in the Bayesian inference.

Key Results

  • High Success Rate: The proposed EnBIP system achieved a 90% success rate on the real-world bimanual insertion task, even with varying starting positions.
  • Superiority over Baselines: It significantly outperformed standard methods like Behavior Cloning (0% success) and Probabilistic Motor Primitives (ProMP) with Time Warping (20% success), primarily because those methods struggle to adapt to the "stop-and-go" nature of physical obstructions.
  • Robustness to Disturbance: In simulations, the robot successfully completed tasks even when researchers manually pushed the arm backward, as the phase estimation correctly "rewound" the task progress.

Limitations and Future Work

The primary limitation is the Expertise Requirement. The current model assumes that the human demonstrations are "near-optimal." If a teacher provides a clumsy or incorrect demonstration, the Bayesian model may struggle to find a precise phase estimate.

The authors suggest that future work should focus on making the system more robust to "naive" or non-expert teachers, potentially by filtering out bad data or using reinforcement learning to refine the human-provided primitives.

Summary updated

Abstract

In this paper, we discuss a framework for teaching bimanual manipulation tasks by imitation. To this end, we present a system and algorithms for learning compliant and contact-rich robot behavior from human demonstrations. The presented system combines insights from admittance control and machine learning to extract control policies that can (a) recover from and adapt to a variety of disturbances in time and space, while also (b) effectively leveraging physical contact with the environment. We demonstrate the effectiveness of our approach using a real-world insertion task involving multiple simultaneous contacts between a manipulated object and insertion pegs. We also investigate efficient means of collecting training data for such bimanual settings. To this end, we conduct a human-subject study and analyze the effort and mental demand as reported by the users. Our experiments show that, while harder to provide, the additional force/torque information available in teleoperated demonstrations is crucial for phase estimation and task success. Ultimately, force/torque data substantially improves manipulation robustness, resulting in a 90% success rate in a multipoint insertion task. Code and videos can be found at https://bimanualmanipulation.com/
BibTeX
@inproceedings{stepputtis2022system,
  author = {Simon Stepputtis and Maryam Bandari and Stefan Schaal and Heni Ben Amor},
  title = {{A System for Imitation Learning of Contact-Rich Bimanual Manipulation Policies}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2022},
  url = {https://openreview.net/forum?id=RGmmRytT77},
}

Loading related content...

Figure