DexHand Lab pro: event-hardened tactile dexterity benchmark

[ducthuykh1009]

Registration UUID: 2555924c-74a4-4788-be61-1f1e65bf3f44

Project: DexHand Lab pro
Submission folder: submissions/dexhand_lab

Summary
This PR submits the event-hardened DexHand Lab dexterous manipulation benchmark. The project is a hand-only MuJoCo submission centered on a human-like five-finger robotic hand, blind tactile active perception, tactile pose estimation, adaptive regrasp, precision assembly, 224-degree cap rotation, 9x load hold, tactile combination lock manipulation, no-crush vial uncap/sample delivery, microsuture threading, stress evaluation, and a judge-readable evidence pack.

Latest update in this PR

• Removed stale score/PR backup files from the submitted folder so judge evidence does not point to old PRs or old leaderboard state.
• Replaced hardcoded PR target metadata with current-branch upstream PR metadata in outputs/submission_readiness_report.json.
• Added validator event-hygiene gates for summary/readiness agreement, stale PR target detection, deprecated backup detection, SRT timing inside the generated demo duration, and embedded stress-eval summary evidence.
• Kept the generated demo in the event-required 1-3 minute window: duration_s = 85.25.
• Preserved all existing DexHand evidence: cap rotation, blind tactile mode, tactile pose assembly, combination lock, vial task, microsuture task, contact-causality audit, stress eval, hardware replay audit, and 45-gate task suite.

How to run

python submissions/dexhand_lab/run_demo.py
python submissions/dexhand_lab/run_demo.py --episodes 3 --seed 42 --no-video --difficulty medium
python submissions/dexhand_lab/run_stress_eval.py --seeds 32
python submissions/dexhand_lab/validate_submission.py

Validation run completed locally

• python -m py_compile submissions/dexhand_lab/run_demo.py submissions/dexhand_lab/validate_submission.py submissions/dexhand_lab/quality_gate.py submissions/dexhand_lab/run_stress_eval.py submissions/dexhand_lab/arena_task_suite.py
• python -m unittest discover -s submissions/dexhand_lab/tests -p "test_*.py"
• python submissions/dexhand_lab/run_demo.py
• python submissions/dexhand_lab/run_stress_eval.py --seeds 32
• python submissions/dexhand_lab/arena_task_suite.py
• python submissions/dexhand_lab/microsuture_benchmark.py
• python submissions/dexhand_lab/run_demo.py --episodes 3 --seed 42 --no-video --difficulty medium
• python submissions/dexhand_lab/validate_submission.py

Current evidence highlights

• validation_passed: true
• submission_readiness_pass: true
• rules_alignment_pass: true
• rubric_readiness_pass: true
• task gates: 45/45
• object_snap_events: 0
• contact_causality_pass: true
• cap_rotation_success: true
• cap_rotation_achieved_deg: 224.0
• load_hold_success: true
• load_hold_x: 9.0
• feedback_success_rate: 1.0
• baseline_success_rate: 0.59375
• microsuture_threading_success: true
• demo_video_duration_rule_pass: true

Files to inspect first

• submissions/dexhand_lab/media/demo.mp4
• submissions/dexhand_lab/media/keyframes.png
• submissions/dexhand_lab/outputs/summary.json
• submissions/dexhand_lab/outputs/validator_report.json
• submissions/dexhand_lab/outputs/submission_readiness_report.json
• submissions/dexhand_lab/outputs/event_rules_report.json
• submissions/dexhand_lab/JUDGE_BRIEF.md
• submissions/dexhand_lab/EVIDENCE_INDEX.md
• submissions/dexhand_lab/dataset/task_suite_report.json
• submissions/dexhand_lab/dataset/contact_causality_report.json

Honest scope
DexHand Lab uses simulation-native MuJoCo/contact evidence and deterministic contact-aware controllers. It does not claim learned RL, real camera perception, perfect contact physics, or physical hardware execution. Hardware-related files are replay/audit artifacts for transfer analysis, not a real hardware trial.

[ducthuykh1009]

Update pushed on codex/dexhand-event-final-polish.

Latest cleanup:

• Refreshed PR_DESCRIPTION.md so it matches the current DexHand evidence: 85.25s demo, 45/45 gates, validator/readiness/rules pass, and current task list.
• Removed stale PR-number literals from the validator scan path while keeping stale-target protection active.
• Confirmed no stale PR links, old score text, encoding artifacts, or copied-source references remain in the submission scan.

Validation rerun:

• python -m py_compile submissions/dexhand_lab/validate_submission.py submissions/dexhand_lab/run_demo.py
• python submissions/dexhand_lab/validate_submission.py

Current final evidence remains: validator pass, submission readiness pass, rules alignment pass, 45/45 gates, object_snap_events=0, contact_causality_pass=true, cap_rotation_success=true, load_hold_success=true, microsuture_threading_success=true.