T-RO Revision

First appeared: July 9, 2026
Last seen: July 30, 2026
Appeared on: 18 days

2026

July 2026

check what remaining parts are misaligned with the old-env
run experiment to compare teacher policy with reproduction baseline

Jul 29Wednesday

teacher policy pipeline with polygons is built and able to learn to achieve success.
- Recover custom TensorBoard logs
- Disable noisy output
- Check that the peg and hole is aligned
- Fix the initialization error
- Ensure the gripper starts with a width smaller than the peg diameter
- Fix the polygon centering issue by aligning the peg’s true center with the hole center, then compensating the hand pose
- Achieve successful insertion using the descending diagnosis script
- Zero out the reward term related to force minimization, since it is not the goal
- Recap:
- At 3200 iter with 0.5 success rate threshold successful, it progresses curriculum to 1.0667
- The keypoint not being straight up is intentional. keypoint_radius can tilt the keypoints to make yaw visible in the reward

done add failure taxonomy and match horizon length
resolve the final migration for multiasset, reward, and no-gui
fixing basic config and logging and enable baseline reproduction
inspect the results of claim1-report.txt
plan migration plan for Chunk H and I, teacher policy and student encoder
implementing teacher policy in isaaclab

review shapes, reward, and termination migration results
get minimal contour shape ablation results. Haven't inspect it

Get claim 5 results and continue migration

Jul 22Wednesday

running Claim 5 experiments (3D vs. 6D action)
get contour shape-aware teacher policy exp values, log in Code Tasks P0
log what new asset-changing mechanism fails and why
minimal contour shape ablation, run exp

implementation simple contour-aware teacher policy on old env, Code Tasks P0
run contour shape-aware teacher policy exp on old env, Code Tasks P0
continue isaaclab migration

begin policy and PF-only reproduction baseline on old env. (change from VPF to PF-only due to available disk size issue)
log policy and PF-only metrics (log at Code Tasks P2)
create migration map in chunk
begin migration, see migration-map and progress-and-next-steps on server

fix IsaacLab RTC (Raycast snippet !isaac-ts)
get soft peg-insertion env launching on IsaacLab

skim Pre-registration again. looks good.
trace old IsaacLab code and setup coding agent

Jul 15Wednesday

Finish AutoMate (Tang et al., NVIDIA, RSS 2024), closed.
討論維護 Paper Reading List 的權責

Finish AutoMate (Tang et al., NVIDIA, RSS 2024) Pass 2, Pass 3
After practice Time-box Reading Guide once, the process seems to be fine

reply the Hamaya san with the intended direction
revise again Pre-registration
NVIDIA’s AutoMate trains insertion/assembly on hundreds of geometry. Watch out this when making claims.
Setup Paper Reading List, which AI mentor should help me to maintain it
Try time-box reading on AutoMate (Tang et al., NVIDIA, RSS 2024)

1st, 2nd version of Pre-registration

identify the majors issue: overclaim generalizability, shape-aware policy, need to justify the real-world test, 6D action.
discuss and decide issues priorities with AI mentor.
working on Pre-registration, but it is difficult to write.

現在面對 T-RO review 沒有那麼緊張了
review-3 processed.

T-RO review-1, review-2, processed.