Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
video
video
label
class label
0 classes
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null

SmolVLA Top-Down Grasp Evaluation

Eval of yianW/smolvla_mug_grasp_topdown.

How the eval was built

Each episode uses the exact scene parameters (object_euler, object_scale, pedestal_height, xy_offset) from a successful training trajectory (from rerot_jitter_results.json, jitter_from_successes_results.json, and jitter_round2_results.json). Up to 5 diverse configs per (mug, base_rot) case, evenly sampled across the z-rotation range. One subprocess per episode so the Genesis morph is baked with the correct per-episode euler (previous eval reused one scene per mug and applied set_quat, which composes with the baked morph rotation and therefore didn't produce the intended world orientation).

Overall

13/52 = 25.0% success (lift object z > 3 cm within 150 steps).

Per-case results

Case Success
cup_rot000 2/5 (40%)
cup_rot180 5/5 (100%)
mug_2_tripo_rot180 2/5 (40%)
mug_5_tripo_rot090 1/5 (20%)
mug_7_tripo_rot180 2/5 (40%)
mug_9_tripo_rot000 0/5 (0%)
mug_9_tripo_rot270 0/5 (0%)
mug_dowan_rot090 0/5 (0%)
paper_coffee_cup_rot090 1/5 (20%)
paper_coffee_cup_rot270 0/2 (0%)
paper_cup_20_oz_rot270 0/5 (0%)

Layout

Per-trial folders <case>_e<ep_idx>/ each contain episode.mp4 (third-person video, 20 fps) and result.json (scene params + success + z_delta + step count). all_results.json is the concatenated list.

Task

Grasp mug from top and lift

Downloads last month
239