Robo-Diffusion: Sampling Algorithm Design

Objective

Design better sampling algorithms for diffusion models that achieve high quality while maintaining efficiency.

What You Can Modify

• Sampling algorithm type (DDPM, DDIM, DPM-Solver, custom)
• Number of sampling steps
• Sampling strategy parameters (eta for DDIM, order for DPM-Solver)
• Adaptive sampling strategies
• Custom sampling schedules

What Is Fixed

• Network architecture (task-specific standard architecture)
• Diffusion model training
• Evaluation environments

Evaluation

Evaluated on three D4RL MuJoCo environments:

1. hopper-medium-v2
2. walker2d-medium-v2
3. halfcheetah-medium-v2

Metrics: normalized_score, episode_reward, inference_time, sampling_steps

Baselines

ddpm

DDPM sampling with 100 steps - Standard but slow

ddim

DDIM sampling with 20 steps - Faster deterministic sampling

dpm_solver

DPM-Solver++ with 10 steps - Fastest high-quality sampling