Hugging Face's logo

MMMMoK9922
/
MetaRoboBench-Speed-Control

Reinforcement Learning
PyTorch
robotics
rainbow-dqn
speed-control
metarobobench
Model card Files
xet
 Community
MetaRoboBench-Speed-Control / MODEL_MANIFEST.md
MMMMoK9922's picture
MMMMoK9922
Update public handoff status
2804651 verified
preview code
|
Raw History Blame Contribute Delete
16 kB

MetaRobo Hugging Face Model Handoff

Inventory date: 2026-08-21

Source of truth: dev-machine and the shared vePFS archive. A800-CCI is no longer available and is not a recovery source.

Hugging Face account verified on dev-machine: MMMMoK9922.

Published public repositories:

  • MMMMoK9922/MetaRoboBench-Pi05
  • MMMMoK9922/MetaRoboBench-Speed-Control
  • MMMMoK9922/MetaRoboBench-Flow-DPO

Upload Status

Upload status at 2026-08-21 13:30 Asia/Shanghai:

  • MetaRoboBench-Speed-Control: complete, 99 files and 0.324 GiB.
  • MetaRoboBench-Pi05: 21650, frame-aligned 21060, and normal-taskframe20 10550 complete; remaining checkpoints are queued.
  • MetaRoboBench-Flow-DPO: all three complete models are uploaded.
  • The repositories were changed from private to public after the private-storage quota was reached.
  • No train_state files were uploaded.
  • A fresh public-LFS probe succeeded after the visibility change. The remaining Pi05 uploads resumed automatically.
  • Pending: 10 Pi05 inference checkpoints at the time of this update.

Upload Rules

  • Upload each Pi05 checkpoint's params/, assets/, and _CHECKPOINT_METADATA while preserving the directory layout.
  • Do not upload train_state/ in the default archive. The 13 Pi05 train states add about 391.22 GiB and are only needed for exact optimizer-state resume.
  • Upload DPO params/ together with metrics.jsonl and the corresponding launcher/config.
  • Upload lightweight speed/value heads together with summary.json and their runtime configuration.
  • The repositories are public by explicit handoff decision. Confirm base-model and dataset licensing before redistributing weights beyond this archive.
  • Never upload AK/SK, Hugging Face tokens, proxy values, task platform credentials, logs containing secrets, caches, or raw rollout data.

Minimum Must-Preserve Set

This is the first upload batch if time is limited. Approximate payload: 92.7 GiB plus small metadata.

Priority HF path Source Size Reason
P0 pi05/ff500_chunk30/21650 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_metarobo_clean_speed_balanced2500_adaln_cfgdrop_v2speed070130/metarobo_clean_speed_balanced2500_ff500_chunk30_adaln_cfgdrop_bs256_w16_10ep_save1_keep5_20260722_092158/21650 11.59 GiB params Main available speed-control base. Clean Fast/Normal/Slow: 73.0/75.4/78.0.
P0 pi05/frame_aligned/21060 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_metarobo_clean_speed_taskbalanced_framealigned2500_adaln_cfgdrop_20260731/taskbalanced_framealigned_ns17_nf21_fs6_ss6_ff30_bs256_w32_10ep_save2_keep2_20260731/21060 11.59 GiB params Final task-balanced/frame-aligned SFT and DPO reference.
P0 pi05/normal_taskframe20/10550 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_metarobo_clean_normal_taskframe20_balanced2500_adaln_cfgdrop_20260804/normal_taskframe20_s68_fp12_ff20_bs512_w32_10ep_save1_keep2_20260804/10550 11.59 GiB params Normal-speed, 20% false-frame model; important failure-ratio ablation.
P0 pi05/robotwin2_h50/30000 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_robotwin2_piper_video_37task_clean_h50_base_20260817/robotwin2_piper_video_37task_clean_h50_base_bs256_lr5e5_g8_s30000_save10k_keep2_20260817/30000 11.59 GiB params Final 37-task RoboTwin2 Piper H50 base checkpoint.
P0 pi05/robotwin2_merged/40000 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_robotwin2_piper_video_merged_base_20260817/robotwin2_piper_video_base_e10_bs256_lr5e5_g8_s118315_save10k_keep2_20260817/40000 11.58 GiB params Latest available merged RoboTwin2 Piper base checkpoint.
P0 rainbow/shared_all50_final/shared_all50.pt /kairos_vepfs_volc/data/manlichen/MetaRoboBench/speed_rl/p05ff500c30_vlmrainbow50_shared_g32_20260820/models/shared_all50.pt 4.45 MiB Final shared 50-task Rainbow head: 1941/2500 = 77.64%, 100,491 decisions, 99,962 updates.
P0 flow_dpo/g16_s500_conservative /kairos_vepfs_volc/autodrive/manlichen/data/MetaRoboBench/flow_dpo_runs/pi05_flow_dpo807_g16_s500_conservative_20260805 11.59 GiB params Conservative 500-step DPO model; completed metrics through step 500.
P0 flow_dpo/g4_s3000_b01_a005 /kairos_vepfs_volc/autodrive/manlichen/data/MetaRoboBench/flow_dpo_runs/pi05_flow_dpo807_g4_s3000_b01_a005_20260804 11.59 GiB params Completed 3,000-step DPO model.
P0 flow_dpo/g8_s1000_b01_a005_r2 /kairos_vepfs_volc/autodrive/manlichen/data/MetaRoboBench/flow_dpo_runs/pi05_flow_dpo807_g8_s1000_b01_a005_r2_20260803 11.59 GiB params Completed 1,000-step DPO model.

Complete Pi05 Research Archive

The 13 available Pi05 inference parameter trees total 150.65 GiB. Upload all entries below for a complete research handoff.

Priority HF path Source Size Purpose/status
P0 pi05/ff500_chunk30/21650 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_metarobo_clean_speed_balanced2500_adaln_cfgdrop_v2speed070130/metarobo_clean_speed_balanced2500_ff500_chunk30_adaln_cfgdrop_bs256_w16_10ep_save1_keep5_20260722_092158/21650 11.59 GiB Best available speed-control development base.
P0 pi05/frame_aligned/21060 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_metarobo_clean_speed_taskbalanced_framealigned2500_adaln_cfgdrop_20260731/taskbalanced_framealigned_ns17_nf21_fs6_ss6_ff30_bs256_w32_10ep_save2_keep2_20260731/21060 11.59 GiB Final frame-aligned checkpoint. Clean F/N/S: 69.0/65.6/76.6.
P0 pi05/frame_aligned/16848 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_metarobo_clean_speed_taskbalanced_framealigned2500_adaln_cfgdrop_20260731/taskbalanced_framealigned_ns17_nf21_fs6_ss6_ff30_bs256_w32_10ep_save2_keep2_20260731/16848 11.59 GiB Earlier frame-aligned checkpoint. Clean F/N/S: 67.4/65.6/75.8.
P0 pi05/normal_taskframe20/10550 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_metarobo_clean_normal_taskframe20_balanced2500_adaln_cfgdrop_20260804/normal_taskframe20_s68_fp12_ff20_bs512_w32_10ep_save1_keep2_20260804/10550 11.59 GiB Normal-only, approximately 20% false-frame experiment.
P1 pi05/natural_failure_raw/30110 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_metarobo_clean_speed_balanced2500_adaln_cfgdrop_v2speed070130/metarobo_clean_speed_balanced2500_natural_failure500_adaln_cfgdrop_bs256_w16_10ep_save1_keep5_20260718_015133/30110 11.59 GiB Raw natural-failure-500 ablation. Clean F/N/S: 60.8/59.0/62.2.
P1 pi05/natural_failure_taskmeantrim/10910 /kairos_vepfs_volc/data/manlichen/MetaRoboBench/checkpoints/pi05/pi05_metarobo_clean_speed_balanced2500_adaln_cfgdrop_v2speed070130/metarobo_clean_speed_balanced2500_natural_failure500_taskmeantrim_ptsfix_r1_adaln_cfgdrop_bs512_w16_16gpu_10ep_save1_keep5_20260720/10910 11.59 GiB Task-mean-trim and ptsfix natural-failure model. Clean F/N/S: 62.4/63.1/64.4 (partial result).
P1 pi05/speed_equal/10890 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_metarobo_clean_speed_equal2500_adaln_cfgdrop_v2speed070130/metarobo_clean_speed_equal2500_adaln_cfgdrop_bs512_w16_16gpu_10ep_save1_keep5_20260711_r1/10890 11.59 GiB Equal-speed data ablation. Clean F/N/S: 69.0/71.2/72.2.
P1 pi05/fail30/21690 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_metarobo_clean_speed_fail30_balanced2500_adaln_v2speed070130/metarobo_clean_speed_fail30_balanced2500_adaln_v2speed070130_bs256_w16_10ep_20260709/21690 11.59 GiB Older approximately 30% failure-frame/data-ratio ablation.
P1 pi05/speed_balanced_retrain/20000 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_metarobo_clean_speed_balanced2500_adaln_cfgdrop_v2speed070130/metarobo_clean_speed_balanced2500_adaln_cfgdrop_bs512_w16_16gpu_10ep_save1_keep2_20260810_r1/20000 11.59 GiB Retrained speed-balanced mainline replacement candidate.
P0 pi05/robotwin2_h50/30000 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_robotwin2_piper_video_37task_clean_h50_base_20260817/robotwin2_piper_video_37task_clean_h50_base_bs256_lr5e5_g8_s30000_save10k_keep2_20260817/30000 11.59 GiB Final 37-task RoboTwin2 Piper H50 checkpoint.
P1 pi05/robotwin2_h50/20000 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_robotwin2_piper_video_37task_clean_h50_base_20260817/robotwin2_piper_video_37task_clean_h50_base_bs256_lr5e5_g8_s30000_save10k_keep2_20260817/20000 11.59 GiB Earlier 37-task RoboTwin2 Piper H50 checkpoint.
P0 pi05/robotwin2_merged/40000 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_robotwin2_piper_video_merged_base_20260817/robotwin2_piper_video_base_e10_bs256_lr5e5_g8_s118315_save10k_keep2_20260817/40000 11.58 GiB Latest merged RoboTwin2 Piper checkpoint.
P1 pi05/robotwin2_merged/30000 /kairos_vepfs_volc/autodrive/manlichen/workplace/MetaRoboBench-RoboTwin/policy/pi05/checkpoints/pi05_robotwin2_piper_video_merged_base_20260817/robotwin2_piper_video_base_e10_bs256_lr5e5_g8_s118315_save10k_keep2_20260817/30000 11.58 GiB Earlier merged RoboTwin2 Piper checkpoint.

Speed And Value Heads

These files are small and should all be uploaded with their result summaries. The final shared Rainbow head depends on Pi05 21650.

Priority HF path Source Payload Notes
P0 rainbow/shared_all50_final /kairos_vepfs_volc/data/manlichen/MetaRoboBench/speed_rl/p05ff500c30_vlmrainbow50_shared_g32_20260820 4.45 MiB model plus summary Final shared model: 77.64% over 2,500 episodes; Fast/Normal/Slow chunks 38.6/32.3/29.1%; about 4.23 switches/episode.
P1 rainbow/shared_all50_accel_checkpoint /kairos_vepfs_volc/data/manlichen/MetaRoboBench/speed_rl/p05ff500c30_vlmrainbow50_shared_g32_accel_resume_20260820 4.45 MiB Distinct earlier/accelerated checkpoint: 16,500 decisions and 15,964 updates. Its SHA-256 differs from the final model.
P1 rainbow/per_task_all50 /kairos_vepfs_volc/data/manlichen/MetaRoboBench/speed_rl/p05ff500c30_vlmrainbow50_full_g16_resume_20260819 222.65 MiB across 50 heads plus summary Per-task Rainbow collection and fixed-speed/adaptive comparisons.
P2 rainbow/vlm_pilot8 /kairos_vepfs_volc/data/manlichen/MetaRoboBench/speed_rl/p05ff500c30_vlmrainbow8_pilot_g8_20260818 35.62 MiB across 8 heads Early VLM Rainbow pilot.
P2 rainbow/pilot8 /kairos_vepfs_volc/data/manlichen/MetaRoboBench/speed_rl/p05ff500c30_rainbow8_pilot_g8_20260817 20.71 MiB across 8 heads Earlier non-final pilot.
P1 cql/all50_v1 /kairos_vepfs_volc/data/manlichen/MetaRoboBench/speed_rl/p05ff500c30_all50_cql_v1_g8_20260814 242 KiB model plus summary All-task CQL baseline.
P1 cql/v3_sweep /kairos_vepfs_volc/data/manlichen/MetaRoboBench/speed_rl/p05ff500c30_cql_v3_g8_20260814 590 KiB across 3 heads plus summary CQL coefficient sweep.
P1 candidate_value/qslow_active_ensemble /kairos_vepfs_volc/data/manlichen/MetaRoboBench/candidate_value/p05ff500c30_qslow_active_g8_20260813 28.43 MiB across 8 heads plus summary Active Q-slow candidate-value ensemble. Accuracy was not sufficient for the final controller, but it is a key negative result.
P1 candidate_value/qslow_vlm_ensemble /kairos_vepfs_volc/data/manlichen/MetaRoboBench/candidate_value/p05ff500c30_qslow_vlm_g4_20260813 14.16 MiB across 4 heads plus summary VLM candidate-value baseline; collapsed to selecting Fast in the recorded test.

The RECAP/risk/hazard directories do not currently contain a formal trained risk checkpoint suitable for publication. They contain source code, reports, feature files, manifests, or preflight cache files. Preserve their code in Git, but do not label them as trained models in Hugging Face.

DPO Archive

The three complete DPO parameter trees total 34.77 GiB.

Priority HF path Source Status
P0 flow_dpo/g16_s500_conservative /kairos_vepfs_volc/autodrive/manlichen/data/MetaRoboBench/flow_dpo_runs/pi05_flow_dpo807_g16_s500_conservative_20260805 Complete 500-step conservative run.
P0 flow_dpo/g4_s3000_b01_a005 /kairos_vepfs_volc/autodrive/manlichen/data/MetaRoboBench/flow_dpo_runs/pi05_flow_dpo807_g4_s3000_b01_a005_20260804 Complete 3,000-step run.
P0 flow_dpo/g8_s1000_b01_a005_r2 /kairos_vepfs_volc/autodrive/manlichen/data/MetaRoboBench/flow_dpo_runs/pi05_flow_dpo807_g8_s1000_b01_a005_r2_20260803 Complete 1,000-step run.

Do not upload the following as model checkpoints:

  • pi05_flow_dpo807_g8_s20000_b01_a005_20260803: metrics reached step 6,292, but no full params/ tree remains.
  • pi05_strict_flow_dpo_pilot49_20260803_v3: empty/incomplete metrics and no full model.
  • pi05_strict_flow_dpo_pilot49_20260803_v3_16gpu: empty/incomplete metrics and no full model.

Missing Critical Asset

The historical Quality-Mix AdaRMS mainline checkpoint 10850 is missing.

  • It was the strongest historical baseline after inference-dropout correction: Clean Fast/Normal/Slow 72.0/75.8/77.8, Random 50.2/56.4/56.6.
  • Its original checkpoint directory is absent from dev-machine and the vePFS checkpoint archive.
  • Historical logs indicate checkpoint janitor cleanup removed it.
  • A800-CCI is no longer available, so it cannot be used as a recovery source.
  • Upload the configs, launchers, and evaluation summaries that reference 10850, but do not create a placeholder model or imply that the weight is available.

Repository And Reproducibility Metadata

Attach the following to every model card or repository README:

  • Code branch: codex/vlm-rainbow-shared-20260821
  • Code commit: 5b36567f0c6174ad82a3d4c669134f7b28acc945
  • GitHub: https://github.com/MMMM3202/MetaRoboBench/tree/codex/vlm-rainbow-shared-20260821
  • Evaluation protocol: 30 submitted actions, natural JAX policy noise, inference metadata dropout disabled unless explicitly marked otherwise.
  • Training dataset identifier, metadata proportions, normalization/stats source, training steps, and any known incomplete evaluation fields.

Upload Order

  1. Create the three private Hugging Face repositories and verify a small-file upload.
  2. Upload Pi05 21650, then final shared Rainbow, then frame-aligned 21060.
  3. Upload all three complete DPO models.
  4. Upload 10550, RoboTwin2 H50 30000, and merged 40000.
  5. Upload the remaining Pi05 ablations and earlier checkpoints.
  6. Upload per-task Rainbow, CQL, and candidate-value negative-result heads.
  7. Generate SHA-256 manifests from the source paths and verify the remote file inventory before considering the handoff complete.

Resume After Quota Increase

The resumable scripts and state files are stored on dev-machine at:

/kairos_vepfs_volc/autodrive/manlichen/offboarding_20260815/hf_upload

After increasing the account's private storage quota, run:

cd /Users/manlichen/kaifaji
offboarding_20260815/hf_upload/resume_after_quota.sh

The scripts skip completed checkpoints and continue from the first pending model. They retain a 15-minute per-command timeout and automatic retry protection for stalled proxy connections.