Reproduce “Robotic Object Insertion with a Soft Wrist through Sim-to-Real Privileged Training” on OpenResearch

About this paper

Yuni Fuchioka, Cristian C. Beltran-Hernandez, Hai Nguyen, Masashi Hamaya

OMRON SINIC X Corporation researchers introduced a sim-to-real privileged training framework for compliant robotic object insertion using a soft wrist. The method achieved high success rates, including 100% for circular peg insertion with zero misalignment, and demonstrated generalization to unseen square peg shapes on a physical robot without real-world training, reducing engineering efforts for complex assembly.

Key insights

• The sim-to-real privileged training framework effectively addresses the dual challenges of soft robot partial observability and the real-world data burden for RL, enabling autonomous control from proprioceptive sensors.
• Decomposing the learning problem into teacher-student stages (privileged training) significantly boosts learning efficiency and success rates compared to direct unprivileged RL for complex manipulation.
• Domain randomization across key parameters like grasp angle, hole position, and controller gains is essential for generating policies robust enough to achieve zero-shot transfer to a physical robot and handle real-world uncertainties.

Summary by alphaXiv.