How Well do Diffusion Policies Learn Kinematic Constraint Manifolds?
Lexi Foland, Thomas Cohn, Adam Wei, Nicholas Ezra Pfaff, Boyuan Chen, Russ Tedrake
AI summary
Problem
It remains unclear how well diffusion policies inherently learn and respect implicit kinematic constraints in robot data, independent of confounding factors like low-level controllers or gripper compliance.
Approach
The authors collect teleoperation data for a constrained bimanual pick-and-place task, systematically perturb it to create datasets with varying constraint violations, train diffusion policies on these datasets, and evaluate both task success and constraint adherence in simulation and on real hardware.
Key results
- Diffusion policies learn only coarse approximations of kinematic constraint manifolds
- Policy performance and constraint adherence degrade with smaller dataset sizes
- Training on noisier data with intentional constraint violations reduces success rates
- Manifold curvature shows inconclusive correlation with constraint satisfaction and task success
Why it matters
Provides critical insights for robotics researchers and practitioners deploying diffusion policies for constrained manipulation tasks, highlighting the need for high-quality data and realistic constraint evaluation beyond task success metrics.
Abstract
Diffusion policies have shown impressive results in robot imitation learning, even for tasks that require satisfaction of kinematic equality constraints. However, task performance alone is not a reliable indicator of the policy’s ability to precisely learn constraints in the training data. To investigate, we analyze how well diffusion policies discover these manifolds with a case study on a bimanual pick-and-place task that encourages fulfillment of a kinematic constraint for success. We study how three factors affect trained policies: dataset size, dataset quality, and manifold curvature. Our experiments show diffusion policies learn a coarse approximation of the constraint manifold with learning affected negatively by decreases in both dataset size and quality. However, manifold curvature showed inconclusive correlations with constraint satisfaction and task success. A hardware evaluation verifies the applicability of our results in the real world. Project website with additional results and visuals: https://diffusion-learns-kinematic. github.io/. This work was supported by Amazon.com, PO No. 2D-15694048, the Office of Naval Research, PO No. N000142412603, SRI International, PO No. PO81455, the Defense Science and Technology Agency, PO No. DST00OECI20300823, the NSERC CGS D-587703, and the National Science Foundation Graduate Research Fellowship Program under Grant No. 2141064. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation (or other funding organizations). The authors are with the Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology, Cambridge, Massachusetts. Corresponding author: lkfoland@mit.edu