Shiki42/robotwin_blocks_ranking_rgb_hybrid_200_dynFcam

robotwinblocksrankingrgbhybrid200dynFcam

A validated LeRobot v2.1 release of two native RoboTwin 2.0 expert schedules for blocks_ranking_rgb. The variants use independent accepted seeds and are concatenated into one training split; matching episode offsets are not paired scenes.

dynFcam observation derivative

This repository reuses exactly the same validated native episodes as `Shiki42/robotwin_blocks_ranking_rgb_hybrid_200` at revision 18fb86f7571012b7b2c3b7340aa57cf8573b783a. It was rebuilt without running RoboTwin simulation: for every aligned sample, cam_high is selected from the corresponding native head_camera frame, while both wrist cameras remain native per-frame observations. State, action, phase, arm mask, seed, scene identity, and episode ordering are unchanged.

The _dynFcam suffix means dynamic front/main camera. The legacy source dataset intentionally routes most ASYNC main-camera samples to segment-start scene context (with a dynamic boundary window); this derivative removes that routing for a controlled visual-observation ablation.

Composition

Total: 200 episodes, 97,125 aligned samples, 600 videos, 30 FPS. The name hybrid_200 means a 100+100 union of two schedule variants under the same physics/camera configuration; it does not mix incompatible configurations. Exact output episode, source episode, seed, scene identity, and source revision mappings are in parallelvla_variant_manifest.json.

What is the official RoboTwin schedule?

The official blocks_ranking_rgb schedule is classified as `mixed_schedule_unqualified`, not as either release variant:

• —most grasp, lift, place, retreat, and color-order primitives are serial;
• —when the active arm changes, the official script may pair the next block grasp with the opposite arm's back_to_origin;
• —the audited official clean-50 HDF5 has aggregate active-motion overlap ratio 0.1269146608.

A single paired move does not make the complete schedule dependency-aware. The official clean-50 archive was not copied into this release because it also lacks source commit, recoverable actor-pose scene identity, phase trace, native collision/replay receipts, synchronized training-action replay receipts, and the required (1024, 6) point cloud.

Schedule semantics

non_overlapping_reference

Every grasp, lift, place, retreat, and back_to_origin primitive is sent through a native single-arm planner move. At most one arm has active joint motion at a time. Aggregate active-motion overlap ratio is 0.

dependency_aware_overlap

The color order and final red-green-blue ranking constraints remain unchanged. When consecutive blocks use different arms, a ready grasp may be paired with the previous arm's dependency-safe back_to_origin only after the preceding place and retreat dependencies are satisfied. The paired primitive is executed by native RoboTwin move(left_action, right_action). Aggregate active-motion overlap ratio is 0.1245310130; median active overlap ratio is 0.0909090909.

There is no random temporal interpolation, offline HDF5 retiming, frame deletion/duplication, or observation splicing. The _dynFcam derivative uses the native main-camera frame at every aligned sample and does not synthesize or freeze waits.

Source revisions and configuration

• —parallelVLA _dynFcam derivative commit: f1e99f77302c4c87e6c34e35c8b7fd586bc13c2d

• —parallelVLA native generation commit: ce20e57befbd4c424bb70f094aa19b4066d43f99
• —parallelVLA LeRobot publication adapter commit: ffc8b27fdd86af8668739cafcc75396838dc57d6
• —RoboTwin native generation commit: c3ddfa8b97d5519efa828b075999bd0006778e5e
• —phase-aware upstream LeRobot converter checkout: fa416dd14cd004d8d005cbeaa3a960299d46a78a
• —base physics/camera config: parallelvla_clean_pcd
• —raw HDF5 provenance schema: parallelvla_schema_version=4
• —final seed policy: independent

Training alignment

The conversion is checked element-by-element against the same shared preprocessing used by training and replay validation:

• —preprocessor: robotwin_next_sample_joint_target_v2, version 2
• —alignment: state[t] -> action[t+1]
• —deployment replay profile: shared_dp3_pi05_qpos_topp_per_target_v1
• —raw/sample and LeRobot FPS: 30
• —main camera routing: native dynamic frame at every aligned sample
• —wrist camera routing: native dynamic frames per arm
• —prompts are tagged per frame with [PHASE=SYNC] or [PHASE=ASYNC]

Features include 14-DoF observation.state, 14-DoF action, observation.phase_one_hot, observation.arm_active_mask, and head/left-wrist/right-wrist AV1 video at 640x480.

Validation

Both variants pass 100/100 native plan, task-success, native replay, artifact, and inter-arm collision gates.

• —native inter-arm contact steps: 0 for both variants
• —dependency_aware_overlap synchronized training-action replay: 100/100 passed
• —training-action replay inter-arm contact steps: 0
• —all 200 processed episodes match the shared action preprocessor element-by-element
• —all 600 LeRobot videos strictly decode with matching episode frame counts; no repair was required
• —all 200 processed episodes retain element-exact state/action/phase/mask arrays; 1,800 sampled camera frames exactly match their native source-frame conversion
• —dynamic-camera validation receipt: validation/dynamic_camera_validation.json

Validation reports are included under validation/.

Limitations

This LeRobot release does not include raw (1024, 6) point clouds or full native phase/collision receipts; those remain in the validated RoboTwin roots. Because the variants use independent seeds, schedule efficiency must be compared on a separate common-seed diagnostic set, not by pairing LeRobot episode indices.

Citation

Please cite RoboTwin 2.0 and parallelVLA when using this release.