cloudwalk-research/psi0-g1-sneaker-6ep-v2-eval

Psi0 G1 Sneaker-in-Box — 6 episodes (v2 held-out eval)

Held-out evaluation set for the v2 Psi0 fine-tune. 6 episodes (3 LEFT + 3 RIGHT) split off from the canonical 205-episode v2 source and intentionally never seen during training. The training partner is psi0-g1-sneaker-199ep-v2; together they reconstruct the canonical source exactly (6 + 199 = 205). Used in the v2 Phase 11.5 open-loop eval (2026-05-07) — see Open-loop eval results below.

TL;DR. v2 memorized rather than generalized. At ckpt_40000 the OOD/ID gap is 5.5×–7.2× on hand and arm joints — ID is healthy (the inference path is sane), OOD blows past the runbook's predicted thresholds by 5×/8×. Training past 30 k steps made it worse on OOD arm_joints (regression 0.609 → 0.628 → 0.632) while ID arm_joints kept improving — that's the textbook overfitting signature, and it makes ckpt_30000 the Pareto-better deploy candidate. The model is also 3× worse picking with the RIGHT hand (OOD hand_joints 0.214 vs LEFT 0.070 rad) despite the training set being only 1.19× LEFT-heavy (108 L / 91 R) — so RIGHT-side scene diversity (object pose, lighting, distractors) is the bottleneck, not raw episode count. Verdict: Phase 14 (real-robot eval grid) is gated; if shipping is forced, prefer ckpt_30000.

Biggest lesson for v3 (single highest-leverage intervention): wire the OOD probe (this dataset) into the inner training loop as an early-stopping signal. Today Psi0's eval/loss is in-distribution by default (val_repo_ids = train_repo_ids); the OOD curve is invisible during training. Without that probe, we trained blind past the 30 k OOD optimum.

Quick facts

Episodes	6
Frames	1 791 @ 30 fps
Format	LeRobot v2.1
Camera	egocentric, 640×480, H.264 yuv420p
Robot	Unitree G1 + Inspire RH56DFTP (5-finger, 6 active DoF/side, padded to 14-D)
Hand-side composition	3 LEFT (eval idx 3, 4, 5) + 3 RIGHT (eval idx 0, 1, 2)
Action dim	36
State dim	32
Task instruction	"Pick up the sneaker and place it inside the box."
Derived from	psi0-g1-sneaker-205ep-v2-source — episodes [41, 56, 91, 101, 162, 181]
License	Apache-2.0

meta/stats.json and meta/stats_psi0.json are recomputed over these 6 episodes (not copied from the canonical source); they describe what's actually in this set, not training-time normalization stats — those stay frozen on the model side.

The 6 episodes

eval idx	orig idx	side	margin	length (frames)	left_unused	right_unused
0	41	RIGHT	0.5275	216	1.000	0.472
1	56	RIGHT	0.5610	295	1.000	0.439
2	91	RIGHT	0.5314	255	1.000	0.469
3	101	LEFT	0.5453	276	0.455	1.000
4	162	LEFT	0.5425	424	0.458	1.000
5	181	LEFT	0.5211	325	0.479	1.000

meta/episodes.jsonl carries an extra orig_episode_index field per entry, mapping each eval index back to its position in the canonical source.

Selection criteria

Hand-side is computed from action means on each canonical-source episode:

left_unused  = mean(action[:, 0:6])    # ≈1.0 = LEFT idle
right_unused = mean(action[:, 7:13])   # ≈1.0 = RIGHT idle
side   = "LEFT" if left_unused < right_unused else "RIGHT"
margin = abs(left_unused - right_unused)

All 205 source episodes are unambiguously L or R (min margin 0.27). The 6 held-out episodes were chosen for high margin (> 0.52), spread across the canonical episode_index range per side (RIGHT idx 0–93, LEFT idx 94–204) to capture cross-day collection variation, and varied lengths (216–424 frames, 1.96× span). Total held-out frames (1 791) = 2.87 % of the canonical 62 447.

How to download

import os
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='cloudwalk-research/psi0-g1-sneaker-6ep-v2-eval',
    repo_type='dataset',
    local_dir='data/real/psi0-g1-sneaker-6ep-v2-eval',
)

Run scripts/data/patch_lerobot_meta.py after download. For the eval invocations that produced the results below, see Reproducing this.

Open-loop eval results (v2 fine-tune, 2026-05-07)

Setup

Phase 11.5 ran the v2 fine-tune (cloudwalk-research/psi0-g1-sneaker-199ep-v2-finetune, ckpts 20 k / 30 k / 40 k) through scripts/eval/openloop_eval_psi0.py four times: ID = 6 random episodes from psi0-g1-sneaker-199ep-v2 (training set, seed 42 → eps [17, 85, 86, 128, 150, 170], 1 771 frames); OOD all-6 = all 6 episodes of this dataset (1 791 frames); OOD-RIGHT = --episode-indices 0 1 2 (766 frames); OOD-LEFT = --episode-indices 3 4 5 (1 025 frames). Common: --stride 1 --num-inference-steps 10, single GPU (RTX 5090). Training-time normalization stats are preserved across all 4 invocations — only the dataloader source changes. Errors are denormalized L1 norm per dim group, smaller is better.

Figure 1 — Glance view: each polygon is one checkpoint; bigger polygon area = better overall on that metric. Score per axis is 1 − (error / max_error_in_panel), so the outer edge = perfect (zero error) and the origin = worst observed. hand_joints panel: ckpt_30k (teal) and ckpt_40k (red) have nearly identical areas, both substantially larger than ckpt_20k (purple) — the latter collapses at the OOD-RIGHT axis, dragging its area down. arm_joints panel: every polygon collapses into a thin vertical spike pointing only at the ID axis (top) — all three checkpoints have near-zero scores on every OOD axis, visualizing that the model fails at OOD arm equally regardless of training step. Radial signature: hand-side ckpt selection matters; arm-side ckpt selection does not, because no ckpt generalizes there.

Figure 2 — Open-loop convergence on 4 evals × 3 checkpoints. ID (blue, train-v2 random 6) keeps improving 20k → 40k on both metrics; OOD all-6 (red, this dataset) plateaus on hand_joints and regresses on arm_joints past 30k; OOD-LEFT (green dashed, idx 3,4,5) tracks OOD-all closely while OOD-RIGHT (orange dashed, idx 0,1,2) sits visibly above on hand_joints. The growing ID-vs-OOD divergence IS the overfitting signature; ckpt_30000 sits at the OOD optimum.

Figure 3 — ckpt_40000 snapshot. ID (blue) is tiny next to the OOD bars — that's the memorization gap. hand_joints: OOD-RIGHT (0.213) is 3.06× OOD-LEFT (0.070) despite the training set being only 1.19× LEFT-heavy (108 L / 91 R). arm_joints: LEFT and RIGHT are within 7% (0.661 vs 0.617). The asymmetry is hand-specific (fingertip-trajectory / grip-timing), not reach-specific — RIGHT-side scene diversity is the bottleneck for v3.

Figure 4 — Cross-checkpoint magnitude comparison (Figure 2 as bars, side-by-side). Reads top-down per ckpt: hand_joints ID is in monotone freefall (0.042 → 0.017 → 0.014); OOD all-6 (red) is stuck at ~0.10 from 30k onward; OOD-RIGHT (orange) bar towers above all others at every ckpt (0.231 / 0.212 / 0.213) and is the dominant contributor to the OOD-all bar. arm_joints ID drops 0.205 → 0.137 → 0.114; the three OOD bars are tightly clustered around 0.61–0.66, with a slight upward creep from 30k → 40k. The hand-side asymmetry is the only place where ckpt selection moves the needle on OOD.

Figure 5 — Asymmetry evolution across training. The RIGHT-vs-LEFT hand_joints multiplier grows during training: 2.23× at 20k → 3.11× at 30k → 3.06× at 40k. The 30k step is where the LEFT-side error drops sharply (0.103 → 0.068, down 34%) while the RIGHT-side barely moves (0.231 → 0.212, down 8%) — extra training disproportionately benefits the better-covered side, widening the asymmetry rather than closing it. By 40k the asymmetry has plateaued and arm_joints has started regressing on both sides — late training is past the OOD optimum.

Table A — ID vs OOD vs per-side @ ckpt_40000

Group	ID (train-v2 random 6)	OOD all-6	OOD-LEFT (idx 3,4,5)	OOD-RIGHT (idx 0,1,2)	OOD/ID ratio	OOD–ID delta	per-side |L−R|/max
hand_joints	0.0138	0.1000	0.0698	0.2135	7.2×	+0.0862	67.3% (R 3.06× worse)
arm_joints	0.1140	0.6318	0.6614	0.6168	5.5×	+0.5178	6.7% (symmetric)
height (m)	6.7e-6	6.3e-6	6.0e-6	7.0e-6	~1.0×	~0	—
rpy / vx / vy / torso_vyaw / target_yaw	0.0	0.0	0.0	0.0	—	0	— (constant in dataset)

hand_joints OOD is 7.2× the ID baseline — the model fits training-distribution finger trajectories cleanly (ID 0.0138 rad ≈ v1's 0.0117) but generalizes weakly to held-out demos. arm_joints OOD is 5.5× the ID baseline; the absolute (0.632 rad ≈ 36°) is large enough that real-robot reach trajectories will look qualitatively different from the demo. The remaining rows (height, rpy, vx, vy, torso_vyaw, target_yaw) are constants in the training distribution and reproduce to zero error at inference — nothing to learn, nothing to generalize.

Table B — Per-checkpoint OOD all-6 (the headline convergence signal)

ckpt	hand_joints	arm_joints
20 000	0.1201	0.6090 ← best
30 000	0.1009	0.6278
40 000	0.1000 ← best	0.6318

ID at the same ckpts: hand 0.0417 → 0.0173 → 0.0138 (monotone, 3.0×); arm 0.2045 → 0.1369 → 0.1140 (monotone, 1.8×). OOD hand barely moves (1.20×); OOD arm regresses past 30 k. The ID-vs-OOD divergence (Figure 2) is the overfitting signature — extra training buys ID accuracy at the cost of OOD generalization.

Table C — Per-checkpoint per-side OOD breakdown

ckpt	LEFT hand	LEFT arm	RIGHT hand	RIGHT arm
20 000	0.1034	0.6189	0.2308	0.6167
30 000	0.0683	0.6480	0.2125	0.6121
40 000	0.0698	0.6614	0.2135	0.6168

RIGHT hand is 3× LEFT at every ckpt; arms symmetric within ~5 %. The bottleneck is fingertip-trajectory / grip-timing generalization on RIGHT-pickup scenes — not RIGHT-side reach. Train-v2 is 108 L / 91 R (1.19× LEFT-heavy), too small to explain a 3× error gap by itself → RIGHT-side scene diversity (object pose, lighting, distractors) is the primary lever for v3, raw episode count secondary.

Decision

ID anchor is healthy (hand 0.0138 rad ≈ v1's 0.0117; arm 0.114 rad ≈ 2× v1's 0.058, attributed to bimanual mirroring widening the arm distribution) — the inference path is sane. But OOD blows past the runbook's predicted thresholds:

• hand_joints ≤ 0.020 rad → actual 0.100 rad (5× over)
• arm_joints ≤ 0.080 rad → actual 0.632 rad (8× over)

Verdict: do NOT proceed to Phase 14 (real-robot eval grid) on ckpt_40000 without further investigation. If a real-robot test must happen, ckpt_30000 is the better candidate (best OOD-arm at 0.628 vs 0.632; hand parity with 40 k). Expect coarser end-effector tracking on RIGHT-pickup scenes.

v3 priorities, in order of expected impact: (1) RIGHT-side scene diversity (sneaker pose, lighting, distractors — episode-count alone won't close the 3× hand gap), (2) recovery / negative scenes (carry-over from v1, never addressed in v2), (3) OOD probe wired into the inner training loop as an early-stopping signal (today eval/loss is in-distribution by default — we trained blind past the 30 k OOD optimum), (4) higher episode count (~1.5–2×) to keep effective-epochs lower at comparable step count. See the model card for the deployer-facing version.

Reproducing this

EVAL_RUN_DIR=".runs/finetune/sneaker-ext-left-real-eval"
COMMON=(--run-dir "$EVAL_RUN_DIR" --ckpt-steps 20000 30000 40000 \
        --stride 1 --num-inference-steps 10)

# ID — random 6 episodes from training set (seed 42)
python scripts/eval/openloop_eval_psi0.py "${COMMON[@]}" \
    --num-episodes 6 --output openloop_id_v2.json

# OOD — all 6 held-out episodes
python scripts/eval/openloop_eval_psi0.py "${COMMON[@]}" \
    --eval-dataset-id cloudwalk-research/psi0-g1-sneaker-6ep-v2-eval \
    --num-episodes 6 --output openloop_ood_v2.json

# OOD-RIGHT subset
python scripts/eval/openloop_eval_psi0.py "${COMMON[@]}" \
    --eval-dataset-id cloudwalk-research/psi0-g1-sneaker-6ep-v2-eval \
    --episode-indices 0 1 2 --output openloop_ood_right.json

# OOD-LEFT subset
python scripts/eval/openloop_eval_psi0.py "${COMMON[@]}" \
    --eval-dataset-id cloudwalk-research/psi0-g1-sneaker-6ep-v2-eval \
    --episode-indices 3 4 5 --output openloop_ood_left.json

Wrapper: scripts/eval/run_v2_id_ood_evals.sh runs all four sequentially (~60–90 min on 1× RTX 5090).

Raw artefacts. JSON reports and training curves (W&B run 3yifcdgs) tracked internally at the CloudWalk Robotics Lab; the per-checkpoint per-side numbers reproduced above are the same that feed Tables A–C.

Related datasets

Repo	Episodes	Hand split	Purpose
psi0-g1-sneaker-94ep-v1	94	0 L / 94 R	v1 fine-tune (right-hand only)
psi0-g1-sneaker-205ep-v2-source	205	111 L / 94 R	canonical immutable source
psi0-g1-sneaker-199ep-v2	199	108 L / 91 R	v2 training partner of this dataset
psi0-g1-sneaker-6ep-v2-eval (this)	6	3 L / 3 R	v2 held-out open-loop eval

Attribution

Developed by cloudwalk-research in the CloudWalk Robotics Lab (CW-RL). Part of the Psi0 collection. Held-out probe for the Psi0 base model; stored in the LeRobot v2.1 format.

Citation

@misc{cwrl_psi0_g1_sneaker_6ep_v2_eval_2026,
  title        = {Psi0 G1 Sneaker-in-Box --- 6 held-out episodes (v2 OOD eval)},
  author       = {{CloudWalk Robotics Lab}},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/cloudwalk-research/psi0-g1-sneaker-6ep-v2-eval}
}

@article{psi0_2026,
  title         = {Psi0: A Generalist Vision-Language-Action Model for Humanoid Robots},
  author        = {Wei, Songlin and others},
  year          = {2026},
  eprint        = {2603.12263},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2603.12263}
}

@article{black2024pi0,
  title   = {{$\pi_0$}: A Vision-Language-Action Flow Model for General Robot Control},
  author  = {Black, Kevin and Brown, Noah and Driess, Danny and others},
  journal = {arXiv preprint arXiv:2410.24164},
  year    = {2024}
}

@software{lerobot2024,
  title  = {{LeRobot}: State-of-the-art Machine Learning for Real-World Robotics in {PyTorch}},
  author = {{Hugging Face}},
  url    = {https://github.com/huggingface/lerobot},
  year   = {2024}
}