Embedded Files
Qualitive Videos of behavior learned with ASTEROID
Qualitive Videos of behavior learned with ASTEROID
Real-world Robot Manipulation
To test whether ASTEROID can learn policies deployable on real hardware, we learn an in-context exploration policy for a vision-denied robot manipulation task. Here, the robot must locate and pick up a cube randomly initialised (in a 10x2cm grid) using only the proprioception from the robot arm and gripper. Since the cube is unobserved, the policy must explore candidate cube locations, attempt grasps, and use the history of gripper poses and grasp outcomes to infer the cube location. We train the policy entirely in simulation using Omni-Reset (Yin et al.) and then transfer it to the real robot. ASTEROID achieves 60% success rate, while the BC baseline has a 20% success rate!
ASTEROID enables rummaging inside a box to extract objects, as cameras are occluded when searching for objects.
ASTEROID enables rummaging inside a box to extract objects, as cameras are occluded when searching for objects.
ASTEROID can explore without relying on vision inputs, so it enables exploration under visual distractors!
ASTEROID can explore without relying on vision inputs, so it enables exploration under visual distractors!
ASTEROID (6/10 success)
ASTEROID (6/10 success)
Policy explores the space efficiently and is able to pickup the cube after a few attempts.
Policy explores the space efficiently and is able to pickup the cube after a few attempts.
BC (2/10 successes)
BC (2/10 successes)
Policy does not know how to explore and is trained on expert data only. As a result, after the first grasp attempt the history goes out of distribution the policy fails. (Note: it succeeds only if the first attempt is successful)
Policy does not know how to explore and is trained on expert data only. As a result, after the first grasp attempt the history goes out of distribution the policy fails. (Note: it succeeds only if the first attempt is successful)
Bonus: As the policy is trained to explore via proprioception only, it generalises to unseen objects and colors.
Rebuttal Ablations
Rebuttal Ablations
Ablation A: Sub-sampling expert history in ASTEROID
Ablation B: OOD generalisation of ASTEROID in the 2D Navigation Env.
Ablation C: Long Horizon Evaluation
RMA baseline
BC+PPO on Pegincreased to 2 seeds
Ablation D: Sub-sampling on KeyDoor Task
ASTEROID (Ours)
ASTEROID (Ours)
RMA
RMA
A. Peg Insertion using wrist cameras leads to efficient exploration and recovery behavior.
ASTEROID (Ours)
ASTEROID (Ours)
BC
BC
B. Exploration using only tactile sensors.
ASTEROID (Ours)
ASTEROID (Ours)
DPT (Lee et al.)
DPT (Lee et al.)
BC i.e. Naive Distillation on expert data fails because of lack of support over contexts and no-exploration.
C. Exploration using only tactile sensors and no proprioception.
D. Exploration in complex procgen navigation environments is learned purely through supervision
E. Exploration Behavior in Habitat.
F. Ant navigation to hidden goals shows targeted exploration in high-dimensional control spaces.
G. ASTEROID enables memory in KeyDoor Task.
Page updated
Google Sites
Report abuse