Main Real World Results (featured by high-speed, fine-grained movements, sophisiticated gaiting, and realiability)
Diverse “Pen” Spinning Skills
Pass-v0
01 Pass-v0 continuous success
02 Pass-v0 continuous rotation (10.5 rotation episodes)
Pass-v1
01 Pass-v1 continuous success
02 Pass-v1 continuous rotation (4 rotation episodes)
Pass-v2
01 Pass-v2 continuous success
02 Pass-v2 continuous rotation (2 rotation episodes)
Pass-v3
01 Pass-v3 continuous success
02 Pass-v3 continuous rotation (3 rotation episodes)
Thumb-Around
01 Thumb-Around (rigid object) task continuous success
04 Diverse Thumb-Around Rotations
Simple Sonic-1
01 Simple Sonic-1 task with 50 in 50 success
04 Sonic Rotations
01 Simple Sonic-1 task with 50 in 50 success
01 Simple Sonic-1 task with 50 in 50 success
Three-finger-pass (palm up)
01 Three-finger-pass (palm up) continuous success
02 Three-finger-pass (palm up) continuous rotation (25+ mins, 91 episodes x 2pi)
01 Three-finger-pass (palm up) continuous success
02 Three-finger-pass (palm up) continuous rotation (25+ mins, 91 episodes x 2pi)
01 Three-finger-pass (palm up) continuous success
02 Three-finger-pass (palm up) continuous rotation (6 mins)
Three-finger-pass (palm down)
01 Three-finger-pass (palm down) continuous success
02 Three-finger-pass (palm down) continuous rotation (13+ mins)
Three-finger-pass-v2 (palm down)
01 Three-finger-pass-v2 (palm down) continuous success
02 Three-finger-pass-v2 (palm down) continuous rotation (5+ mins)
Four-finger-pass (palm down)
01 Four-finger-pass (palm down) continuous rotation (30+ mins, 134 x 2pi in total)
02 Four-finger-pass (palm down) continuous rotation (15+ mins, 134 x 2pi in total)
Four-finger-pass (palm down)
01 Four-finger-pass (palm down, leap) task continuous success (50 in 50)
02 Four-finger-pass (palm down, leap) task continuous success (48 in 50)
03 Four-finger-pass (palm down, leap) task continuous rotation
04 Four-finger-pass (palm down, leap) task continuous rotation
Zero-shot Adaptation to 1) different object materials, 2) changed dynamics, 3) various object shapes.
Zero-shot Adaptation
01 Adapation to new geometry (cylinder -> hexagonal prism, octagonal prism; continuous rotation & multiple successes)
02 Adapation to unseen materials (rigid -> soft; continuous rotation & multiple successe)
03 Adaptation to diverse shapes (T shape, L shape, ellipse, c-ellipse)
04 Adapation to a thinner object (2.0 cm -> 1.4 cm)
05 Adaptation to new object shapes (T shape, L shape, long board)
06 Adaptation to OOD object mass
Some interesting results from our experiments.
Few-shot Adaptation
01 A low-friction object (from lees than 50% success rate to 94% success rate (47/50))
02 Remove the tape wrapped around the fingers (from less than 10% success rate to around 72% success rate)
02 Cuboid -> Octagonal Prism
02 Thumb-around (unssen materials, rigid -> soft; from less than 10% success rate to 96% success rate (48/50))
03 New object geometry (from around 40% to 100% success rate)
04 Object with a larger diamerter (1.4 cm -> 1.6 cm, from around 50% success rate to 100% success rate)
Robustness to Human Disturbances
Robustness to Human Disturbances
01 1) Push the object out of the balanced pose, 2) Drag the object to an extreme initial position
Initial position distribution, rollout distribution
Robustness to initial position distribution, transition stochasticity, and hardware instability
01 Consecutive success one-episode rollout (59/59)
02 Visualization on the rollout distribution
Fun Results
Fun Results
01 Fun Result (1) - Pass-v2
02 Fun Result (2) - Policy resulting from an intermediate reward searched by GPT
The idea
TODO
The method