Symskill: Symbol and Skill Co-Invention for Data-Efficient and Reactive Long-Horizon Manipulation

• Accepted @ ICRA 2026

• Best Paper Award @ CoRL 2025 Learning Effective Abstractions for Planning (LEAP) Workshop

TL;DR: SymSkill jointly learns symbolic predicates, operators, and stable SE(3) DS skills from 5 min of unlabeled play demos, then plans symbolically to compose skills and recover from failures in real time on long‑horizon manipulation tasks.

Video Introduction

Simulation Experiment

Learning Operators and Skills from individual tasks, then combining

Want Even More Data for Training Your Own Policy? Diffusion Policy trained with DS Data Augmentation

In our demos, there is not enough data for Diffusion Policy. Thus, we use the learned DS policy for data augmentation. The orange lines are augmented data. The black line is the DP rollout. Even with data augmentation, DP policies still cannot perform most tasks with any success. 

Online Algorithm of Symskill 

Disturbance Rejection; Reactive Obstacle Avoidance using Modulation (1x Speed)

Demo with operator annotation (1x speed)