DiMSam: Diffusion Models as Samplers for
Task and Motion Planning under Partial Observability
[paper]
Xiaolin Fang1 , Caelan Reed Garrett2, Clemens Eppner2,
Tomás Lozano-Pérez1, Leslie Pack Kaelbling1, Dieter Fox2
MIT CSAIL1 NVIDIA2
Overview
Difficulties in Task and Motion Planning
Continuous constraints
Sampled trajectories reach the goal set, collision-free, IK feasible, etc.
Partial observability
Microwave door occludes the base
Unknown dynamics
Door dynamics depend on an unknown articulation model
Our approach
Diffusion models as parameterized samplers in TAMP
Conditional sampling of diffusion models for different constraints
Sampling in latent space without explicit state estimation or object models
Experiment
TAMP: Joint searching and sampling
Task1 | Close
Initialization
A microwave is located on the table.
An obstacle located between the door and the base of the microwave.
Goal
The microwave is closed.
Final plan
Move(obstacle)
Push(microwave)
Sampled values
Pose_o obstacle placement
Traj_o obstacle moving trajectory
Traj_m microwave moving trajectory
Traj_r robot moving trajectory
Task2 | Stow - Close
Initialization
A cup and a microwave are located on the table.
Goal
The cup is in the microwave.
The microwave is closed.
Final plan
Stow(cup)
Push(microwave)
Sampled values
Pose_o cup placement
Traj_o cup moving trajectory
Traj_m microwave moving trajectory
Traj_r robot moving trajectory
Task3 | Stow - Close - Blocked
Initialization
A microwave is on the table.
A bowl is located near the microwave that may block the door from being fully opened.
Goal
The bowl is in the microwave.
The microwave is closed.
Additional constraint
The microwave needs to be fully opened before stowing the object.
Final plan
Move(bowl)
Push(microwave)
Stow(bowl)
Push(microwave)
Sampled values
Pose_o_1 bowl intermediate placement
Pose_o_2 bowl final placement
Traj_o_1 bowl moving trajectory
Traj_o_2 bowl stowing trajectory
Traj_m_1 microwave opening trajectory
Traj_m_2 microwave closing trajectory
Traj_r_1 robot door opening trajectory
Traj_r_2 robot door closing trajectory
The start of Traj_m_2 must be the same as the end of Traj_m_1 (microwave latent state). The planner searches for the plan skeleton and samples values jointly that satisfy such constraints. Same for Traj_o and Traj_r.
Guided Sampling
Real world deployment
We apply the model trained in simulation directly to the real world without finetuning on door closing and opening tasks. In both tasks, conditional samplers are used. The sampled trajectories are further rejected by IK and motion sampler, with collision checking to the observed partial point cloud.
In the door-opening task, the robot is holding a stick. The predicted waypoints are set as the target location of the stick endpoint.
Despite the partial observability due to 1) self-occlusion, 2) missing depth reading of reflective material, 3) missing depth of the glass door, our model is still able to perform the task. See the attached videos for observed point cloud.
Failure cases
Predicted trajectory falls short. The microwave door is not closed to the end.
Collision due to 1) stick slips in gripper, and 2) sparsity of the point cloud.
@inproceedings{fang2023dimsam,
title={{DiMSam: Diffusion Models as Samplers for Task and Motion Planning under Partial Observability}},
author={Xiaolin Fang and Caelan Reed Garrett and Clemens Eppner and Tomás Lozano-Pérez and Leslie Pack Kaelbling and Dieter Fox},
booktitle={arXiv preprint arXiv:2306.13196},
year={2023},
}