Kavin60606/so101_partner_193

SO-101 Teleop — Partner Length-Normalized (193 episodes)

Training arm P of the ISR vs Partner-sampling SmolVLA experiment. This is the client's SO-101 teleop set processed with the partner company's method — every episode resampled to the dataset mean length and smoothed — for comparison against the raw baseline and the ISR arm.

TL;DR

	value
Source	makermods/200ep_blue_cube_orange_box (SO-101, real teleop)
Method	Length-normalize + smooth — warp each episode to the mean length, smooth actions, blend video
Episodes	193 (holdout [96,97,98,196,197,198] excluded — never trained)
Frames	54,275 → 54,233 (each episode resampled to 281 frames = dataset mean)
fps	30 (uniform grid) · format LeRobot v3.0 · h264 video
State/action	float32[6] = [pan, lift, elbow, wrist_flex, wrist_roll, gripper], degrees
Cameras	observation.images.{front,wrist} 480×640

What the partner method does

Every episode is time-warped to a fixed length (the dataset mean = 281 frames) and smoothed:

• Target length: mean(episode lengths over the 193 train episodes) = 281 frames.
• State / action: linearly interpolated between the two nearest source rows at each warped position, then low-pass smoothed (gaussian_filter1d, σ = 1 frame).
• Video: each output frame is a linear blend of the two nearest source frames ((1−w)·frame_lo + w·frame_hi) — temporal smoothing of the imagery.
• Result re-indexed onto a uniform 30 fps grid.

This DOES interpolate — unlike ISR (which keeps real frames), the partner method synthesizes new frames (blended images, interpolated actions). Episodes become uniform length; pauses are preserved (just rescaled), not removed.

Contrast with ISR (the experiment's point)

	ISR (so101_isr_193)	Partner (this)
Axis of uniformity	information (motion content)	rescaled time (fixed length)
Pauses	removed	kept (rescaled)
Frames	real, selected (59% kept)	interpolated/blended (length-normalized)
Frame count	adaptive per episode	fixed = 281
Interpolation	none	yes (actions + images)

The comparison asks: does information-uniform selection (ISR) beat length-normalization + smooth (partner) — and does either beat the raw baseline — on a SmolVLA policy.

fps / 30 Hz alignment

Stamped at 30 fps on a uniform grid (same as baseline and ISR), so the SmolVLA action-chunk horizon (chunk_size=50) and eval fps match the other arms — the processing method is the only variable.

Split

Holdout [96,97,98,196,197,198] (1,622 frames) is excluded and reserved for evaluation — no arm trains on a holdout frame.

How it was built

data/teleop_std_poc/build_dataset.py --mode partner:

• reads the source via sequential torchcodec range-decode (verified pixel-identical, ~40–90× faster than lerobot's per-frame reader);
• warps to 281 frames (interpolate state/action, blend images, Gaussian-smooth);
• materialises a LeRobot v3.0 h264 video dataset via parallel sharded build + aggregate_datasets merge (~15 min).

Training / experiment

• Policy: SmolVLA (lerobot 0.6.0), byte-identical to the client recipe except the dataset.
• 20,000 steps, batch 16/GPU × 4, bf16.
• Compared against: raw baseline angkul07/mm_SO101_teleop (teleop100) and the ISR arm Kavin60606/so101_isr_193.
• Eval: 6-episode holdout — action MSE/MAE, per-joint MAE, gripper accuracy, CI-MSE.

Status — trained + evaluated

Dataset built + verified (193 eps, 54,233 frames, video). SmolVLA trained 20k steps (Kavin60606/smolvla_so101_partner). Eval on the 6 holdout eps (1622 frames, seed 1000):

metric	baseline teleop100	ISR	partner
action MSE (deg²)	146.83	266.43	149.28
action MAE (deg)	5.840	8.93	5.81
gripper acc	n/a	0.852	0.932
CI-MSE mean	50.28	77.69	46.72
CI-MSE grasp	60.43	155.05	63.92
CI-MSE release	38.09	37.33	29.94

Result: partner method won. Beats the raw baseline on MAE, gripper accuracy, and CI-MSE mean/median/release; ties on global MSE; only marginally worse on grasp (63.9 vs 60.4). Length- normalize + Gaussian smoothing acts as a mild regularizer that preserves the contact phase (pauses kept, just rescaled), so grasp precision survives where ISR's frame-dropping hurt it.