Dataset Viewer
Auto-converted to Parquet Duplicate
contract
string
seed_rule
string
world_size
int64
n_windows
int64
samples
list
samples_sha256
string
one deterministic window per frozen V7 validation recording
20260824 + rank
8
96
[ { "window": 0, "recording_uuid": "03a06ff4-fa57-49fa-a5da-3ff138dafb2c", "master_start": 444, "rank": 0 }, { "window": 1, "recording_uuid": "04278d06-3d4b-452a-b5a9-23009608e5df", "master_start": 510, "rank": 1 }, { "window": 2, "recording_uuid": "07ef7d27-dd3c-49c3-a...
e38481643682a66b298d12d0a6e1d0afbb3b749dae924001bf415ada2ea2bf3d

tacWAM V7 30k evaluation artifacts

This dataset repository contains compact evaluation artifacts for the tacWAM Cosmos3 V7 checkpoint at optimizer step 30,000.

Contents

  • v7_30k/metrics.json: metrics from true from-noise UniPC-20 rollout on one fixed window from each of the 96 V7 validation recordings;
  • v7_30k/eval_sample_manifest.json: exact recording UUIDs, starts, and sample-set checksum;
  • v7_30k/trajectory_manifest.json: identities and window starts for five complete trajectories;
  • v7_30k/v7_trajectory_*.mp4: five GT-versus-prediction trajectory visualizations;
  • v7_30k/v7_trajectory_*.json: sampler, checkpoint, duration, and teacher-forcing sidecars.
  • teleop_smoke/teleop_smoke_two_cam_right_glove_action.mp4: 27.1 s aligned Teleop playback with table/head RGB, right-wrist RGB, right-glove tactile, hand target, and observed arm-state proxy;
  • teleop_smoke/teleop_smoke_two_cam_right_glove_action.json: signal semantics and known alignment defects for that playback.
  • teleop_smoke/teleop_smoke_tujian_taxels_xarm_revo3d.mp4: recommended playback with the original 880-taxel Tujian hand layout, xArm6 3-D kinematics/TCP trail, and command-driven Revo2 3-D hand;
  • teleop_smoke/teleop_smoke_tujian_taxels_xarm_revo3d.json: geometry provenance and numerical tactile/FK validation results.

The trajectory videos are teacher-forced receding-horizon diagnostics: each model call predicts 0.8 s, the first 0.4 s is displayed, and the next call refreshes from a GT observation. They are not autonomous open-loop rollouts. Top-row video/tactile is GT and bottom-row video/tactile is prediction.

The Teleop playback is a data-audit artifact, not a successful policy-training demonstration. Only the hand target is a recorded command. Its arm plot is observed state because the capture omitted the arm command, and its tactile panel is the glove's 15-pad signal rather than robot Revo2 tactile.

The source validation recordings are not redistributed here. This repository also omits the full prediction tensor archive; it retains only the compact metrics and human-viewable outputs.

Full interpretation and the historical V5/V6 comparison are in the tacWAM repository's docs/v7/V7_30K_CHECKPOINT_EVAL_REPORT.md.

Downloads last month
83