starvla-qwenoft-pipette-eerel-crop3

QwenOFT Round 8 — round 7 + wrist_right centre-720 crop (pipette)

Run starvla_qwenoft_pipette_3view_hil293_eerel_state_crop3, trained 2026-09-15 on node G4L3-ZD1-LAX5-000, 3x B200 (GPUs 4-6), planned 20,100 steps, SIGTERM'd after the step-19000 save/eval (~19.2k) to free the GPUs for round 9. A STRICT SINGLE-VARIABLE experiment against round 7 (../starvla_qwenoft_pipette_3view_hil293_eerel_state_crop, see its README): action space, state, data, pause filter and LR are round 7 verbatim — ONLY the wrist_right input geometry moved. Effective batch is 126 (3 procs x batch 21 x accum 2) vs round 7's 128, and 20,100 steps was chosen to match round 7's sample budget (126 x 20,100 ≈ 2.533M ≈ 128 x 19,800).

Round 7's post-mortem had flagged wrist_right as near-idle (blanking it cost ≤4% on any axis) and proposed centre-720 as the "best which-well view" candidate: the whole rack stays in frame at 1.5x magnification. This round tests exactly that. (Round 7's README had provisionally designated "round 8" as the loss-side y fix; that slot was taken by this input-side test instead, and the series then pivoted to the 8-D wrist-pose action space as round 9.)

The change: all three views cropped

• Token count identical to rounds 6/7 (every view still lands at 256x256).
• DataConfig unitree_g1_pipette_3view_oft_ee_rel_crop3; dataset dirs and stats cache SHARED with rounds 6/7 (crop is images-only) — no rebuild.

Results

Selection criterion: raw eval/mse_holdout (eval every 500, save every 1000; baseline mse_zero_action = 0.09193 on this run's holdout subset).

• Best eval: step 2500 — 0.05802 (ratio 0.631). No such file exists (save_interval 1000) — the same between-saves near-miss as rounds 7 (valley at 2000-2500) and 6.
• Best SAVED: steps_3000 — 0.05948 (ratio 0.647). Next-best saved: 5000 (0.06036), 7000 (0.06039); the unsaved 4500 eval (0.05915) also beat every saved step but 3000.
• Per-axis rmse ratio (model/hold-still, normalised units) at 2500: x 0.85, y 0.92, z 0.73; at 3000: x 0.82, y 0.96, z 0.75. Same shape as round 7: x and z well below parity, y barely below.
• Curve shape: sharp descent to a valley at 2500-4500 (0.058-0.0595), then a slow drift up to a 0.062-0.064 plateau from ~10k onward; last eval 19000 = 0.06303. Like round 7, the minimum sits near ONE data pass (2500-3000 x 126 ≈ 315-378k vs 253.6k surviving chunks) and never returns.
• CROSS-ROUND CAVEAT: this run evaluated with 3 ranks, so its fixed holdout subset (and baseline 0.09193) differs from round 7's 8-rank subset (baseline 0.090210). Raw values and ratios are only coarsely comparable across rounds; within-run rankings are exact. On that coarse read, centre-720 on wrist_right neither helped nor hurt the headline (0.631 vs round 7's 0.634 best-eval ratio).
• No round-8-specific probes (per-view ablation, aim slopes) were run — round 9 (8-D wrist-pose action space, ..._hil293plus21_wrist8rel_ state_crop3) superseded this line first.

Checkpoints

Node (~/Projects/starVLA/outputs/.../checkpoints/): steps_1000 … steps_19000 every 1000; no final_model (killed before 20,100). Log /data/jren313/starvla_outputs/oft_round8.log.

Local, md5-verified both ends:

Serving

• Bridge lane pipette_ee_rel_crop3_starvla, profile starvla_oft_pipette_ee_rel_crop3_inspire, checkpoint var STARVLA_PIPETTE_EE_REL_CROP3_CKPT (VLAPolicyBridge).
• Profile view_crop adds wrist_right center:720 on top of round 7's rgb center_square + wrist_left center:448. Server contract unchanged (serve_view_preprocess: resize 270 -> centre crop 256).
• The local config.yaml + config.full.yaml here are patched to base_vlm: Qwen/Qwen3-VL-4B-Instruct (the node copy points at a node-local model path) — re-pulling configs from the node CLOBBERS this patch. Offline probes must load config.full.yaml, not the slim config.yaml (the slim dump drops action_mode_state_map / include_state).

Successor

Round 9 starvla_qwenoft_pipette_3view_hil293plus21_wrist8rel_state_crop3: same three crops, action space widened to 8-D chunk-relative wrist pose (xyz + axis-angle) + 2 thumb dims, data hil293 ∪ plus21.