Search is not available for this dataset
video video 4.63 6.47 | label class label 2
classes |
|---|---|
14train_h5 | |
14train_h5 | |
14train_h5 | |
14train_h5 | |
14train_h5 | |
14train_h5 | |
14train_h5 | |
14train_h5 | |
15train_inspect | |
15train_inspect | |
15train_inspect | |
15train_inspect | |
15train_inspect | |
15train_inspect | |
15train_inspect | |
15train_inspect |
CI-MSE ↔ closed-loop success — experiment artifacts
Everything needed to reproduce the correlation study without re-running training or simulation.
| folder | contents |
|---|---|
demos/train_h5, demos/val_h5 |
raw ManiSkill traj.h5 (obs images, qpos, actions, env states, success flags) + oracle videos + gen_stats.json |
intervals/ground_truth.json |
critical intervals from privileged sim state (grasp, place), 30 Hz frame indices |
intervals/gemini_2.5_pro.json (+ raw responses) |
VLM annotation via ci_mse prompt (express backend) |
intervals/frames_val.json |
global val frame indices dumped (intervals ± 8) |
closed_loop_eval/<ckpt>/eval.json |
30 episodes/ckpt, seeds 200000+, n_action_steps=30; per-episode success / grasped / placed / success step; 3 videos each |
closed_loop_eval_nas10/ |
early evals with n_action_steps=10 (stall behaviour, half the SR) |
offline_predictions_intervals/<ckpt>/predictions.h5 |
[T,30,6] predicted vs target action chunks (degrees), episode_index, frame_local, proprio |
offline_predictions_allframes/ |
stride-2 all-frame dumps for the raw-MSE baseline |
reports/ |
correlation tables (CSV/JSON), scatter plots, segment-wise analysis |
media/ |
dataset inspection: contact sheets, coverage map, joint traces |
Related: datasets Kavin60606/so100_pickcube_train / _val (LeRobot v3.0), checkpoints Kavin60606/cimse-so100-molmoact2-ckpts.
Code + notes: fd-studio/eval/sim_so100/ (DATA.md).
- Downloads last month
- 45