Parv-09/phi_so101_8bin_v1_trim

phi_so101_8bin_v1 — opening-pause trim table

Companion to BrutalCaesar/phi_so101_8bin_v1.

This is not a dataset. It is a 119-row table plus the script that produced it. The original dataset is unmodified and remains authoritative. Applying this table excludes each episode's pre-teleop dead air as a chunk start point, without deleting a single frame from disk.

Why

Every episode begins with the arm sitting still while the operator has not yet moved the leader. Measured: 8,722 of 72,518 frames (12.0%). Left-side episodes averaged 3.53 s of dead air (72-158 frames), right-side 1.38 s (3-103 frames).

Those frames are a problem for a single-timestep policy like ACT. The arm is stationary and the scene is static, so every dead frame is a near-identical observation, yet the recorded action differs depending on how long the operator happened to wait. Worse, at inference the policy has no clock: it predicts "hold still", nothing moves, so the next observation is the same, so it predicts "hold still" again. That is an absorbing state and the arm can freeze for the whole rollout. With chunk_size=50 (50 frames) the risk is concentrated on the left side, where the shortest pause is 72 frames.

How the cut is defined

DROP = the first frame index where any joint has moved more than 1.0° from its start pose:

DROP = argmax_t ( max_j |action[t,j] - action[0,j]| > 1.0 )

Cumulative, so it is immune to per-frame jitter, and because it takes the first crossing it can never extend past the beginning of real motion into an interior pause.

A per-frame movement threshold was tried and rejected. Servo jitter plus the operator's hand resting on the leader exceeds 0.1°/frame from frame 1 in 57 of 119 episodes, so a per-frame test reports a 1-frame pause where the arm has in fact not moved for 100+ frames.

Evidence the cut is safe

Changing the threshold by 20× moves the total cut by only 1.5 percentage points:

threshold	frames dropped
0.5°	8,404 (11.6%)
1.0°	8,722 (12.0%)
2.0°	8,849 (12.2%)
5.0°	9,185 (12.7%)
10.0°	9,471 (13.1%)

The boundary is sharp, so it lands on genuine dead air rather than slow task motion.

What is deliberately NOT trimmed

• Interior pauses. Only 4 of 119 episodes have any interior static run ≥ 0.5 s. Deleting interior frames would be actively harmful: an action chunk is N contiguous frames, so removing interior frames teaches trajectories with teleport jumps no arm can execute.
• Trailing pauses. Present in 117 of 119 episodes (left mean 1.9 s, right 1.2 s). These sit at the end pose with the object already in the box, so they do not recreate the opening absorbing state, and stopping when the task is done is desirable.

How to use it

Trim at the sampler, not the data. Exclude the first DROP frames of each episode as chunk start points:

import csv
from torch.utils.data import DataLoader, SubsetRandomSampler
import torch
from lerobot.datasets.lerobot_dataset import LeRobotDataset

drop = {}
with open("deadtime_per_episode.csv") as f:
    for r in csv.DictReader(f):
        drop[int(r["ep"])] = int(r["DROP"])

ds = LeRobotDataset("BrutalCaesar/phi_so101_8bin_v1", episodes=my_episodes,
                    delta_timestamps=my_delta_timestamps)

froms = list(ds.meta.episodes["dataset_from_index"])
tos   = list(ds.meta.episodes["dataset_to_index"])
a2r   = ds.absolute_to_relative_idx

idx = []
for ep in ds.episodes:
    lo, hi, d = int(froms[ep]), int(tos[ep]), drop[int(ep)]
    assert 0 <= d < (hi - lo)
    idx.extend(a2r[i] for i in range(lo + d, hi))

loader = DataLoader(ds, batch_size=8, num_workers=6,
                    sampler=SubsetRandomSampler(idx, generator=torch.Generator().manual_seed(1000)))

Apply it to both train and validation splits. Untrimmed validation contains ~12% frames whose correct answer is "do not move", which is trivially easy and deflates the reported loss. Trimmed validation reads higher but means something, and is not numerically comparable to untrimmed validation.

deadtime.py regenerates the table from the parquet files and reports the sensitivity analysis above.

Effect on ACT training

Measured on 2 cameras (physical wrist excluded), 80k steps, batch 8, lr 1e-5, seed 1000, chunk 50/75/100. Comparison uses eval-mode scoring on both splits (dropout off, z=0), paired on identical frames.

chunk	untrimmed val L1	trimmed val L1	neutral baseline
50	0.1653	0.1782	~0.189
75	0.2072	0.2067	~0.237
100	0.2389	0.2385	~0.274

The neutral baseline is what validation would read if the trim changed only the exam and not the model (untrimmed best ÷ 0.873, since 12.7% of val frames were removed and scored near zero). All three came in under it, so the model improved on matched-difficulty frames.

Caveats, stated plainly. Single seed, so these margins are not established. And the real purpose of the trim, stopping the arm freezing at episode start, can only be verified on hardware.

Contents

file	what it is
deadtime_per_episode.csv	119 rows: ep, side, frames, DROP, drop_s, drop_pct, kept, trail, trail_s, interior_max, interior_max_s, interior_runs
deadtime.py	the audit script that produced it, including the threshold sensitivity check

Produced for the Φ (Physical Hardware Intelligence) robotics SIG at Northeastern Silicon Valley.