Lift stress

A measured degradation curve, not a victory lap.

Three severity tiers of mass/friction stress on robosuite Lift, paired seeds, Wilson confidence intervals. The baseline passes the gate tier — and visibly breaks under severe stress: in most severe episodes the gripper acquires the cube and then loses the grasp. A verifier that only reports PASS is not credible; the failure cases are the product.

What we tested

Where does a Lift policy break as cube mass and friction move further from nominal?

1.00 → 0.26

success rate, nominal → severe

74%

severe episodes: grasp acquired, then lost

PASS

gate tier (mild) at defined thresholds

Charts

Success vs grasp-failure across severity

rate

Mean peak end-effector force

N (contact proxy)

Metrics

PASS · gate tiersevere score 0.26·50 episodes × 4 conditions, paired seeds

gate: stress success ≥ 80% of nominal
grasp-fail Δ ≤ 15 pp
Wilson 95% CIs on all rates

Metric	nominal	mild	moderate	severe
Success rate	1.00	1.00	0.90	0.26
Success 95% CI	[0.93, 1.00]	[0.93, 1.00]	[0.79, 0.96]	[0.16, 0.40]
Grasp-failure rate	0.00	0.00	0.10	0.74
Mean peak EE force (contact proxy)	7.30	7.82	12.77	14.15
Score (tier / nominal)	—	1.00	0.90	0.26
Verdict at gate thresholds	—	PASS	PASS	FAIL

Methodology

Nominal baseline. Run Lift with default cube mass/friction; record success, grasp-fail, and peak EE force over paired seeds.
Tiered stress. Resample cube mass and sliding friction into mild → moderate → severe ranges while holding seeds paired with the baseline.
Gate + degradation. Score gate tier against success-ratio and grasp-fail Δ thresholds; report severe/nominal as the degradation headline.
Failure logging. Log each failed episode with sampled physics; classify grasp loss with a 5-step debounce after acquisition.

Seeds: 0..49 paired across conditions (50 episodes × 4 conditions)

gate: stress success ≥ 80% of nominal
grasp-fail Δ ≤ 15 pp
Wilson 95% CIs on all rates

Public methodology

Failure case · Grasp slip

The cube gets gripped — then slips. That is the finding.

A Lift policy that scores 1.00 at nominal looks fine until cube mass rises and sliding friction falls. Without severity tiers and an honest grasp-loss metric, this would have stayed a silent green PASS.

How: Paired seeds across mild → moderate → severe mass/friction ranges. Grasp failure = never held, or grasp lost before lift credit (5-step debounce). Each failed episode is logged with the exact sampled physics.

74%

of severe episodes: grasp acquired, then lost (slip)

Seed	Mass (kg)	Sliding μ	Mode
0	2.37	0.082	grasp lost (slip)
1	1.13	0.168	grasp lost (slip)
2	0.90	0.086	grasp lost (slip)
3	2.47	0.282	grasp lost (slip)

Also caught: MuJoCo was ignoring lowered cube friction until cube geom_priority was raised — gripper pads (μ=2.0) otherwise dominated contact. Without that fix the stress test would not have tested what it claimed.

Notes

Sim-only Lift v0.1. MVP proof of a reproducible physics-consistency stress test — not a substitute for real-hardware validation.

Sim-only — validated concept evidence, not a substitute for real-hardware validation or the full platform vision.