We collect teleoperation data for a constrained bimanual pick-and-place task. Then, we perturb these demonstrations to generate three additional datasets that still accomplish the task, but contain increasing constraint violations. We train a policy on each of these datasets and analyze task success and constraint adherence. Lastly, we collect demonstrations for the same task on hardware, train a policy, and evaluate its performance on similar metrics.
Simulation: We analyze the ability of diffusion policies to learn a kinematic constraint between end effectors, evaluating both task success and constraint adherence.
Hardware: We replicate these efforts on hardware, a real-world application.
This shows the success rates of diffusion policies trained on datasets of different sizes and perturbation levels. Each policy was tested over 200 trials.
We evaluate the success of hardware diffusion policies trained on the same task. The top image shows the evaluation distribution, and the bottom shows the hardware task layout.
Here we show how well diffusion policies stick to kinematic constraints. The top tables show the average error in end-effector positions and orientations, while the bottom plots give a sense of how these errors spread across different dataset sizes and perturbations.
Teleoperation data was collected using a specialized teleoperation setup where the operator had explicit control of the self-motion parameter, providing high-quality demonstrations for training.
To examine how constraint violations affect policy performance, data was regenerated with varying levels of perturbations. This allows us to study effects on both task success and constraint adherence.