RSS dual-arm tactile manipulation
This repository holds data for the RSS project. It contains teleoperated demonstrations, and later policy rollouts, recorded on this rig:
- two UFACTORY xArm6 arms, each with a BrainCo Revo2 hand;
- a Tujian tactile glove worn on each robot hand;
- a static head camera and one camera on each wrist.
All new dual-arm data for this project is uploaded here.
Contents
| Folder | What | Notes |
|---|---|---|
task4_dual_arm_tube_pour_20261005/ |
50 successful dual-arm tube-pour teleoperation trajectories, recorded 2026-10-05 | Byte-identical copy of the release in SingleBicycle/tacwam-wetlab-tasks, revision 5bfdc365. Its own README is the data card. Its download examples still name the original repo id. |
task4_dual_arm_tube_pour_fail_cases_20261005/ |
Two failed tube-pour teleoperation trials from the same day: videos (head + both wrists), stills and a README describing each failure | |
task4_dual_arm_tube_pour_20261022/ |
50 more successful dual-arm tube-pour teleoperation trajectories (batch 2), recorded 2026-10-21 and 2026-10-22 with the mesh-based two-arm collision guard | Identical to the folder of the same name in SingleBicycle/tacwam-wetlab-tasks. Its own README is the data card. All 50 are train. |
task4_dual_arm_tube_pour_20261022_failures/ |
7 failed or aborted batch-2 tube-pour attempts with complete recordings, in the same package format | Not for training. Its README gives the label and reason for each. |
task4_pi05_rl_rollouts_20261030/ |
82 real-robot episodes of the RL package: 22 Ο0.5 base-policy episodes (the 20 formal base episodes of 2026-10-30/31 plus 2 acceptance-test episodes), the 39 episodes of RL stage 1 (DSRL-SAC noise-space RL, two learner groups label and label_ours, 10 training + 10 evaluation each; one evaluation slot not run) plus one voided attempt, 20 Diffusion Policy base-policy episodes (dp_base/, 2026-11-01) and the first training episode of the Ο0.5 DSRL-NA stage (pi05_s1_dsrl_na_partial/): three-camera videos, per-decision records, Revo2 and glove tactile, Ours scorer outputs, the learner groups' logs and checkpoints, the stage report, the acceptance-test reports and the reward analysis |
Ο0.5 base policy 13/20; DP base policy 13/20 (3 of its 7 failures are collision-guard aborts). RL stage 1 on Ο0.5 (DSRL-SAC, one round): label 4/10 training β 5/10 evaluation; label_ours 5/10 training β 5/9 evaluation. No measurable difference between the two groups or against the base policy. Its README has the episode tables and the conclusion. |
task4_dp_rl_rollouts_20261103/ |
120 real-robot episodes of the RL package 11-02b with the Diffusion Policy base policy (2026-11-02/03): 20 base episodes, the label_ours_dense group (dense reward = TouchScale scorer progress R + 200 on success; 20 training + 20 evaluation), 20 base-policy evaluation episodes on the same placements, and the label group (binary human label; 20 training + 20 evaluation), plus 2 voided attempts: three-camera videos, per-decision records, Revo2 and glove tactile, Ours scorer outputs, both groups' experience pools, logs and last checkpoints, the two stage reports, the analysis and a README with the conclusions |
Evaluation on the same 20 placements: ours 13/20, label 11/20, base 10/20 (paired McNemar p = 0.51 / 1.0 / 0.73: ours > label > base in direction, not significant at n = 20). Training: ours 17/20, label 15/20. Two ours evaluation episodes ran under another user's CPU load (listed). |
task_usb_plug_dual_20261108/ |
57 successful dual-arm USB-plug teleoperation trajectories (T10: the right hand holds a USB hub, the left hand pinches a flash drive / cable plug and plugs it in), recorded 2026-11-08/09 with preset hand grasps triggered by the glove bend, the fingers-first thumb release and the mesh two-arm guard | Batch 1 of a planned 100 (43 more to come). Its own README is the data card. All 57 are train. Success = the operator's label at the end of each trial, not an independent check. Glove slots are straight in this collection (robot right hand = ITW right slot), unlike the tube-pour batches. |
task_usb_plug_dual_20261108_failures/ |
8 failed USB-plug attempts with complete recordings, same package format | Not for training. Its README gives the label and reviewed reason for each, and lists the trials of the night that are not published. |
task_usb_plug_dual_20261110/ |
43 successful dual-arm USB-plug teleoperation trajectories (T10, batch 2), recorded 2026-11-10 with the teleop push on: in every trajectory the clutch anchor was nudged 5β15 mm before each hand's first grasp and, in most, 3β8 mm while the left hand carried the drive, and the operator corrected; push-ramp frames are flagged action/<side>/arm_perturb_ramp and masked by the loader |
Its own README is the data card. All 43 are train; with batch 1 the 100 planned successes are complete (57 clean + 43 perturb-and-correct). No failed trial in this batch. Glove slots straight, as in batch 1. |
task_usb_plug_dp_base_rollouts_20261115/ |
20 real-robot episodes of the frozen batch-1 USB Diffusion Policy (RL package rl_delivery_usb_20261113c, head camera only, no RL), one per fixed placement P01βP20, 2026-11-15: three-camera videos, every 30 Hz tick, every decision, Revo2 and glove tactile, placement photos |
Base runs: 3/20 success (P05, P07, P09). Its README has the per-placement table and the failure analysis (the policy stalls once both hands hold). The same 20 placements are meant for the later groups. |
task_usb_plug_dp_e600_rollouts_20261117/ |
40 real-robot episodes of four three-camera USB Diffusion Policies trained 600 epochs with different sample weights (demos only, binary filter, vision-reward RA-BC, Ours-reward RA-BC), 10 each on placements P01βP10, 2026-11-17, plus 2 not-scored and 1 voided: three-camera videos, every 30 Hz tick, every decision, Revo2 and glove tactile, placement photos, the operator's stage points | SR 0/10 for all four (no plug fully in). Stage-point score (LingBot-Touch style, 5 stages): Ours 52 %, the other three 42 % each; Ours reached the port in 6/10, the others 2β3/10. Not blind, Ours ran first, 29 of 40 runs stopped early: see its README. |
task_usb_plug_pi05_fix_rollouts_20261118/ |
30 real-robot episodes of three three-camera USB Ο0.5 policies fine-tuned on the 97 teleoperated successes with different sample weights (none, Ours-reward RA-BC, vision-reward RA-BC), 10 each on placements P01βP10, 2026-11-18/19, plus 2 not-counted attempts: three-camera videos, every 30 Hz tick, every decision, Revo2 and glove tactile, placement photos, the operator's stage points, the analysis | SR 0/10 for all three (no plug fully in). Stage-point score (7 steps): plain 35 %, vision 31 %, Ours 27 %; no difference paired by placement (Friedman p = 0.63). Planned 20 each; ended after P01βP10. Not blind: see its README. |
task_usb_plug_pi05_pool2_rollouts_20261119/ |
40 real-robot episodes of four three-camera USB Ο0.5 policies, 10 each on placements P01βP10, 2026-11-19/20: three "pool2" policies trained on the 97 teleoperated successes plus 97 robot / failure episodes (no weights, Ours-reward weights, vision-reward weights) and the demonstrations-only control of 11-18, plus 6 not-counted attempts: three-camera videos, every 30 Hz tick, every decision, Revo2 and glove tactile, placement photos, the operator's stage points, the analysis | SR 0/10 for all four (no plug fully in). Stage-point score (7 steps): control 31 %, pool2 plain 23 %, Ours 20 %, vision 19 %; no pool2 policy above the control (paired p = 0.14β0.34), none differs from another. The control ran last, in one block. Not blind: see its README. |
task_usb_plug_pi05_control3_rollouts_20261120/ |
10 real-robot episodes of the demonstrations-only Ο0.5 control (the 0/1 group) on placements P01βP10, a third evening, 2026-11-20, plus 8 voided attempts: three-camera videos, every 30 Hz tick, every decision, Revo2 and correctly decoded glove tactile, the operator's stage points, the analysis, the glove press tests | SR 0/10. Stage-point score 39 % (the same weights: 35 % on 11-18, 31 % on 11-19/20; Friedman p = 0.38). 8 of 10 stopped because the policy let go early. See its README for the voided attempts and the glove note. |
Tasks
| Task | Status |
|---|---|
| Tube pour: the right hand pours a 50 mL tube into a funnel tube held by the left hand | 100 done (batch 1 + batch 2) |
| RL on the tube pour: a frozen base policy (Ο0.5 on 10-30/31, Diffusion Policy on 11-01β¦03) + noise-space RL (DSRL-SAC), reward = operator label, binary or shaped by the TouchScale scorer | Ο0.5 round (10-31): label 5/10 vs label_ours 5/9 evaluation, no difference. DP round (11-02/03, 120 episodes): evaluation on the same 20 placements ours 13/20, label 11/20, base 10/20 β ours > label > base in direction, not significant at n = 20; training ours 17/20, label 15/20. Both rounds uploaded. |
| USB plug: the right hand holds a USB hub, the left hand pinches a flash drive and plugs it in | 100 done (57 on 2026-11-08/09 + 43 with the teleop push on 2026-11-10) |
| RL on the USB plug: a frozen base policy (the batch-1 Diffusion Policy) on 20 fixed placements; then reward-weighted three-camera Diffusion Policies and Ο0.5 | Base runs (2026-11-15): 3/20. Four 600-epoch DPs (2026-11-17, 10 each on P01βP10): SR 0/10 each; stage-point score Ours 52 %, the others 42 %. Three Ο0.5 (2026-11-18/19, 10 each on P01βP10): SR 0/10 each; stage-point score plain 35 %, vision 31 %, Ours 27 % (7-step rubric). Three Ο0.5 with 97 robot episodes added to the training data (pool2) and the demonstrations-only control (2026-11-19/20, 10 each on P01βP10): SR 0/10 each; control 31 %, pool2 23 / 20 / 19 % (plain / Ours / vision). The same control on a third evening (2026-11-20): 39 %, SR 0/10. Uploaded. Note (2026-11-20): glove data in the robot-rollout folders before 2026-11-20 22:00 are finger-mirrored (see the control3 folder). |
| Pipette: one hand holds a tube, the other transfers liquid with a manual pipette (Finnpipette) | planned |
| Cap: twist a cap off a bottle held by the other hand, and back on | planned |
| Hold and wipe: one hand holds a beaker, the other wipes a mark off its wall | planned |
| Cloth folding, using the task definition of the EgoScale project | planned |
- Downloads last month
- 3,048