Overview
What this challenge is about.
You receive 300 teleoperated demonstrations of a bimanual pour-and-stir task in a Robomimic-style simulator, deliberately including 2 valid solution modes per task (left-pour-right-stir vs. right-pour-left-stir). Train a diffusion policy (DDPM-based action head) on the demos. Train a BC-MLP baseline with identical demos for comparison. Evaluate both on 60 held-out scenarios: success rate, mode coverage (does the policy use both solution modes?), and average completion time. Success is diffusion policy beating BC by at least 15 points on success rate and demonstrating coverage of both modes.
The Brief
What you'll do, and what you'll demonstrate.
Train a diffusion policy on bimanual pour-and-stir demos and prove it handles multimodal solution distributions better than a BC-MLP baseline.
Earning criteria — what you'll demonstrate
- Implement a diffusion-policy action head for imitation
- Design experiments that surface multimodality (mode coverage, not just success)
- Compare diffusion and BC fairly on the same demos
- Write a research note that proposes the next 3 experiments
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
ML Researcher
Implementing diffusion policies and proving multimodality wins is exactly the research-engineering work robot-learning labs hire for in 2024-25.
This challenge sharpens
- diffusion-policies
- imitation-learning
- multimodal-action-distributions
Research Scientist
Designing experiments that surface a method's specific advantage (mode coverage) is the kind of research-design skill industrial research scientists need.
This challenge sharpens
- diffusion-policies
- policy-evaluation
- multimodal-action-distributions
Applied AI Scientist
Translating a research result into 3 concrete follow-up experiments is the applied-AI-scientist craft of moving research toward product.
This challenge sharpens
- imitation-learning
- manipulation
- policy-evaluation