pi05_single — insert_01, step 2000
openpi pi05_single checkpoint for a single-arm, vision-only insertion task.
Step 2000 (1.13 epochs), held-out validation loss 0.1063.
This is the best checkpoint of the run -- validation loss got monotonically worse afterwards.
Configuration
This run differs from the sibling pi05_bi_* models in two ways that make their
loss values not comparable:
pi05_single (this) |
pi05_bi (two_tubes / task2 / bread) |
|
|---|---|---|
| state / action dim | 7 / 10 | bimanual dims |
| image streams used | 1 (camera0 only) |
6 (2 RGB + 4 tactile) |
| lr schedule | constant 2e-4 after 1k warmup | cosine 2.5e-5 -> 2.5e-6 |
The dataset ships six image streams (camera0, camera1, and four tactile
sensors) but pi05_single reads only camera0; camera1 and all tactile inputs
are discarded.
Data
KaiyueChen/insert_01: 500 episodes / 252,795 frames, LeRobot v2.1, 30 fps,
images embedded in the parquet files.
One frame per episode was dropped during preparation. The collection pipeline
writes an all-zero action vector on each episode's terminal frame (it has no
successor action). Because the action vector contains two quaternions whose w
components sit at ~0.99999 with a q01-q99 spread of only 2e-4, quantile
normalization mapped those zeros to about -10,000, and the step-0 training loss
came out at 193,510 while validation read 1.78 (validation always draws the
same leading batches and happened to miss the terminal frames). Dropping the
terminal frame removes 500 of 252,795 frames (0.2%), all of them invalid, and
brings step-0 training loss to 0.74. Merged frame count: 252,295.
Split, held out per source (10%, seed 42): train 450 / val_seen 50 / val_unseen 50.
Normalization statistics (quantile q01/q99) over train only.
Training
| config | pi05_single |
| hardware | 2 x A100-80GB, FSDP |
| batch size | 128 |
| this checkpoint | step 2000 (~1.13 epoch; 1 epoch = 1,764 steps) |
| run length | 6,000 steps (3.40 epochs), 4h08m at 2.4 s/it |
| lr | constant 2e-4 after 1,000 warmup steps |
(pi05_single sets peak_lr = decay_lr = 2e-4, so the cosine schedule does not actually decay) |
|
| LoRA | rank 16 on the LLM, rank 32 on the action expert |
| vision tower | fully fine-tuned -- the freeze filter matches only .*llm.* |
Validation curve
Flow-matching loss, 20 batches per split, same leading batches every evaluation.
| step | train | val_seen | val_unseen | gap |
|---|---|---|---|---|
| 0 | 0.7447 | 0.7284 | 0.7377 | +0.0093 |
| 2000 | 0.0987 | 0.0860 | 0.1063 | +0.0203 |
| 4000 | 0.0881 | 0.0750 | 0.1164 | +0.0414 |
| 5999 | 0.0813 | 0.0673 | 0.1192 | +0.0519 |
val_unseen was already at its minimum by the first validation point (step 2000,
1.13 epochs) and got monotonically worse afterwards, while val_seen improved
monotonically. The gap grew from +0.0203 to +0.0519 (+156%) across three
consecutive points -- an unambiguous overfitting signature, not sampling noise.
For contrast, the sibling six-input pi05_bi runs on this codebase turned much
later and ended with far smaller gaps: two_tubes at ~2.5 epochs (+0.0220), task2
at ~2 epochs (+0.0041), bread at ~2.4 epochs (-0.0037). This run, with a single
camera and no tactile input, turned before 1.1 epochs and ended an order of
magnitude wider. The most likely explanation is the reduced observation space:
with one view and no touch, memorizing training trajectories is an easier way to
lower the loss than learning a transferable policy. pi05_single_vitac
(camera0 + two tactile streams) would be the natural comparison, and insertion
is a task where contact feedback plausibly matters.
Contents
checkpoint/
params/ # inference weights
train_state/ # optimizer state, for resuming
assets/insert_01/
norm_stats.json # computed over the train split only