Curriculum-Aware Diffusion Distillation: Representing Contact Phases for Long-Horizon Door Manipulation

Zhe Han, Bingjie Chen, Zihan Wang, Yizhe Li, Guoping Pan, Yi Cheng, Houde Liu

Keywords: Contact-Rich Manipulation, Diffusion Policy, Behavioral Cloning, Representation Learning

TL;DR: We propose Curriculum-Aware Diffusion Distillation (CADD), converting discrete RL stages into continuous soft labels to modulate diffusion policies for robust long-horizon contact-rich manipulation.

Abstract: Learning continuous control for long-horizon, contact-rich tasks---such as manipulating a heavy articulated door---remains a key challenge in robot learning. While Reinforcement Learning (RL) with privileged state information can solve these tasks in simulation, distilling such policies into vision-based Student networks via Behavioral Cloning often results in catastrophic failure due to state ambiguity and compounding errors. In this paper, we propose Curriculum-Aware Diffusion Distillation (CADD), a framework that bridges this gap through representation design and phase-conditioned generation. CADD explicitly masks privileged states to ensure sim-to-real readiness, maps absolute poses to a continuous Residual 9D action space, and converts hard-coded RL curriculum stages into continuous soft labels. By training a lightweight classification head to predict the current contact phase, CADD modulates a Diffusion Policy via Feature-wise Linear Modulation. Evaluated on the end-to-end door opening task with a challenging bottleneck initialization, CADD achieves a 42.0\% success rate compared to 0\% for Vanilla Diffusion Policies. Qualitative and quantitative evaluations demonstrate that CADD effectively mitigates both the state-ambiguity-induced denoising failure of generative baselines and the spatial compounding errors of deterministic policies at contact boundaries.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 8