pi05_xarm_cube_binmarker_lora
pi0.5 (openpi) LoRA fine-tune for the xArm cube task ("pick up the cube"),
trained on frames with the binmarker overlay: a fixed-length,
fixed-direction arrow per finger (sensor-frame pad normal), shown iff that
finger is in binary contact. NO force magnitude and NO force direction — the
overlay carries only the per-finger contact bit. Ported from
LIBERO_contact_overlay's binbars style.
Training
| |
|---|
| Base | pi05_droid |
| Recipe | LoRA r16/a16 backbone + r32/a32 action expert, attn+ffn |
| Batch / LR | 8 / cosine 1e-4 -> 1e-5, 500 warmup, AdamW, grad-clip 1.0 |
| Data | EdwardoSunny/xarm_cube_binmarker (100 eps, 13103 frames) |
| Early stop | step 7800 |
| Final train loss | 0.0429 |
Variant comparison for this task (identical protocol, only overlay pixels differ):
baseline 0.0437 / points9_arrow_len0 0.0435 / binmarker 0.0429 / points9_arrow 0.0425.
Contact detection (how the threshold was tuned)
Per-finger binary contact from the 3x3x3-axis tactile array:
- Normalization offset = median over known-idle frames (first 10 frames of
each episode, post-homing, gripper open). The naive whole-dataset median is
contact-polluted on tasks that spend most frames in contact (charger: 60%
closed frames made idle read as a fake magnitude of ~1.2).
- Metric = max over the 9 cells of the normalized cell-vector L2 norm
("max-cell"). The cell-sum alternative cannot separate charger idle noise
(p99.5 = 0.45) from real light holds (0.5 on tube/dishwasher's weaker finger).
Max-cell: idle p99.5 <= 0.24 on every task, real contacts >= ~0.3.
- Hysteresis + debounce (Schmitt trigger): ON after 2 consecutive frames
= 0.25, OFF after 3 consecutive frames < 0.18, state reset at episode
boundaries. Light-but-real holds (dips to 0.23) stay latched; single-frame
noise spikes never trigger.
Validated on all 400 episodes (zero idle false positives in every post-homing
window) and visually verified on 5 episodes per task (timelines + frame
montages at approach / contact onset / mid-hold / post-release), plus a final
spot check on the converted LeRobot frames: no arrows during approach, arrows
at contact onset, latched through holds, cleared within <=3 frames of release,
and honest negatives (gripper closed around nothing draws nothing).
Contents
params/ — merged weights (orbax), self-contained for inferenceassets/local/xarm_cube_binmarker/norm_stats.json- Config name:
pi05_xarm_cube_binmarker_lora
Deployment note
At inference, reproduce the SAME overlay on live frames: run the tuned contact
detector (max-cell metric, ON>=0.25 / OFF<0.18, 2-on/3-off debounce, idle-window
offset) and draw the fixed-direction marker (FIXED_DIRECTION, arrow_length_scale=0.12,
thickness 8, dot 22) — see tactile-data-collection/scripts/render_binmarker.py
at commit c64c562 or later (earlier revisions used the measured force direction).
