3D Visual Planning framework to solve long horizon, multi-step rearrangement tasks solely from demonstrations
Submitted to CoRL 2024
- Lean suitable placements for each object
- Retrieve the goal configuration of the task
- Propose multiple placements for each object
- A* search over the suggestions until the goal is reached
(In the initial configuration of the task, the microwave is closed, and the mug is blocking it from opening)
- Trained on 20 demonstrations of the 1 mug version of the task (shown in the figure above).
-The mug placements and door opening angles vary within demonstrations.
- Apply the 3D-VTAMP method, knowing that the microwave is articulated.
- The search can be visualized as a search tree.
- Once a plan is found, plan each robot's motion ahead to find collision-free paths.
- Use the robot model and 3D observations to check for collisions.
Input:
RGBD observation
Names of the objects relevant to the task
Task planning
Task Execution
Input:
RGBD observation
Names of the objects relevant to the task
Task Planning
Task Execution
Input:
RGBD observation
Names of the objects relevant to the task
Task Execution
Greedy Expansion Ablation
This example on the 3 mugs task fails to find the optimal path. It suggests a placement for the red mug that does not move it away from the microwave door opening region
Random Rollouts Ablation
This example of the 3 mugs task also fails. This rollout shows how the selected suggestion for the red mug collides with the green mug in the right.
We trained an end-to-end policy using DP3 with 22 expert demonstrations to finish the task as a baseline. Below shows some of its failure rollouts.
Reached joint limit
Failed to put mug in proper place and grasp the handle
Failed in grasping the mug
Confused two stages (put mug in / open microwave)