Dataset Viewer
Auto-converted to Parquet Duplicate
exact_status
string
exact_status_sha256
string
expert_profile_counts
dict
inventories
dict
source_kind_counts
dict
status
string
totals
dict
unresolved
list
canonical_episodes
list
redactions
list
logs/t4_robotwin_official_aloha_550_exact_status_latest.json
d0d8635f7208af22b0eb4061baf60889d270a826772b601444445746c33b2ea9
{ "final_dogleg080_direct_gap003_primary_d030": 1, "final_dogleg080_direct_gap003_repair_displaced_table_primary_d030": 1, "final_dogleg080_direct_gap003_repair_top_primary_d030": 1, "final_inwardlift040_gap003_primary_d030": 1, "finalarm_offset040_primary_d030": 1, "highlift_primary_d030": 38, "order132_...
{ "logs/t4_official550_inventory_C2.json": "f6699434f1b466f2349fd2073fcdb8649bfb356947186dc2e5014a29f204a696", "logs/t4_official550_inventory_C3.json": "3243bd2f8b35e8dcfb50bd48cb9b1b501ba55be67e30485be1ddcb8e2ca3da43", "logs/t4_official550_inventory_C4.json": "839c83c4a1672eea1d3eea6008829048f4168ba51fd811d25bd0...
{ "alignment_retry": 3, "rescue": 257, "stock": 290 }
PASS_PUBLIC_SANITIZED
{ "canonical_episodes": 550, "duplicate_source_hdf5": 0, "source_bytes": 67546203544, "unresolved": 0 }
[]
[ { "canonical_episode": 0, "expert_profile": "stock", "layout_id": "clean_50/episode_0/seed_1", "max_xy_abs_error_m": 7.170715954352147e-8, "source_episode": 0, "source_hdf5_size": 101903246, "source_kind": "stock" }, { "canonical_episode": 1, "expert_profile": "stock", "l...
[ "source_hdf5", "metadata_path", "machine", "generated_at" ]

TRONCamp Mani T4 exact550

This is a private reproducibility package for the TRONCamp Mani T4 stack_bowls_three experiment. The public code and postmortem live in Datawhale Every Embodied, commit 28c339f.

Contents

  • processed/: complete 550-episode ACT training input, 550 HDF5 files.
  • raw_partial/: local recovery of 368 raw HDF5 files only. The canonical raw manifest expects 550 episodes; 182 raw files are still missing locally and are listed in metadata/raw_partial_manifest.json.
  • checkpoints/: the recovered C3 from-scratch checkpoint and selected exact550 warm-branch checkpoints, with run configs and SHA metadata.
  • metadata/: sanitized dataset identity, official training manifest, processed per-file SHA256 list, raw partial inventory, and checkpoint inventory.
  • repro/: the public scripts, configs, audit tools, and Chinese postmortem.

Reproduction boundary

The processed directory is the complete training input used for ACT. Raw data is included only where it was recovered locally; it is intentionally labeled partial rather than silently presented as complete. Do not publish this dataset until the competition rules, upstream RoboTwin assets, and all third-party data licenses have been checked.

The manifest records 550 canonical layouts, 16-D state/action data, cam_high, cam_left_wrist, and cam_right_wrist, finite values, and the exact next-state alignment contract. The main ACT configuration is ResNet18, encoder 4, decoder 7, hidden 512, FFN 3200, 8 heads, chunk 50, KL weight 10, batch 8, learning rate 1e-5, seed 120, action shift 0, gripper weight 1, and 6000 target epochs.

Integrity

Use the per-file manifests in metadata/ to verify a download before deleting any local copy. The dataset identity records the canonical training-manifest SHA256:

eeca954a7e64dda818a0cca86534e34f6ffa1a8055bd37810f9781a264199901

License and credentials

No access token, private key, machine address, private runtime, or raw experiment logs are included. The repository is private until the data owner confirms the publication boundary.

Downloads last month
96