Robotic Object Insertion with a Soft Wrist through Sim-to-Real Privileged Training

TL;DR

The paper tackles contact-rich object insertion under pose uncertainties using a soft-wrist robot. It introduces a two-stage privileged training framework where a teacher leverages privileged simulation information to learn an insertion policy, and a student encoder learns to estimate this information from onboard sensor histories for real-world deployment, enabling zero-shot sim-to-real transfer. Empirical results show that privileged training improves performance over baselines, that domain randomization enhances robustness, and that the student can accurately recover hidden states such as peg pose and alignment, even for unseen peg shapes. This approach reduces real-world data collection and calibration needs, offering a practical path toward industrial assembly of soft-robot systems.

Abstract

This study addresses contact-rich object insertion tasks under unstructured environments using a robot with a soft wrist, enabling safe contact interactions. For the unstructured environments, we assume that there are uncertainties in object grasp and hole pose and that the soft wrist pose cannot be directly measured. Recent methods employ learning approaches and force/torque sensors for contact localization; however, they require data collection in the real world. This study proposes a sim-to-real approach using a privileged training strategy. This method has two steps. 1) The teacher policy is trained to complete the task with sensor inputs and ground truth privileged information such as the peg pose, and then 2) the student encoder is trained with data produced from teacher policy rollouts to estimate the privileged information from sensor history. We performed sim-to-real experiments under grasp and hole pose uncertainties. This resulted in 100\%, 95\%, and 80\% success rates for circular peg insertion with 0, +5, and -5 degree peg misalignments, respectively, and start positions randomly shifted $\pm$ 10 mm from a default position. Also, we tested the proposed method with a square peg that was never seen during training. Additional simulation evaluations revealed that using the privileged strategy improved success rates compared to training with only simulated sensor data. Our results demonstrate the advantage of using sim-to-real privileged training for soft robots, which has the potential to alleviate human engineering efforts for robotic assembly.

Figures (9)

• Figure 1: We propose a sim-to-real approach for object insertion for a robot with a soft wrist through a privileged training strategy.
• Figure 2: The problem setting considered in this study, showing variable definitions for quantities observable from sensors, versus privileged information only accessible in simulation.
• Figure 3: Overview of the proposed framework. The privileged training has two phases: 1) teacher training: the teacher policy to control the robot is trained with sensor inputs and ground truth privileged information, and 2) student training: the student encoder is trained by running the learned teacher policy to estimate the privileged information from sensor history.
• Figure 4: Pegs' misalignments for the real-robot experiments.
• Figure 5: Snapshots of successful insertion in simulation and real-world. For the simulation, the peg's color turning red indicates that the student encoder detected the alignment of the peg.