Skip to content

Add reaction framework v0.1 with measurable native braking - #4

Open
Tarzelf wants to merge 1 commit into
mainfrom
feat/reaction-framework-v0.1
Open

Tarzelf wants to merge 1 commit into
mainfrom
feat/reaction-framework-v0.1

Conversation

@Tarzelf

@Tarzelf Tarzelf commented Oct 9, 2026

Copy link
Copy Markdown
Contributor

Robotic reaction needs separate evidence for detecting a condition, issuing a brake, actual mechanical stopping, contact behavior, and task outcomes. This adds an offline observation-only monitor and a versioned native linear-actuator/parallel-jaw coupon benchmark. Stop requests continue bounded MuJoCo dynamics to measured sustained rest; they do not freeze the world.

The framework includes timestamped sensor delay/noise/missingness, predictive sphere clearance, contact/slip/grip-loss triggers, latched decisions, matched shadow-monitor/no-brake controls, complete attempted-trial receipts, integrity inspection, and a self-contained HTML report. It reports confusion matrices and uncertainty, sensor-to-actuator timing, stopping path and excursion, force/impulse/penetration, retention, sensing faults, CPU latency/throughput/memory, and optional 1 ms versus 0.5 ms physics comparisons with fixed sensor/reaction cadence. Inspection recomputes measurements and convergence verdicts and refuses omitted attempts or unsupported readiness claims.

Validation: 236 local tests passed with native MuJoCo 3.15.0, installed CLI/dependencies checked, and complete receipts verified. A recorded developmental run has 78 primary attempts plus eight timestep attempts with no runtime errors. In the braking-enabled variant, 24/24 authored positive conditions were detected and 0/15 negatives flagged. Timing agrees within declared timestep tolerances; impact-force/impulse diagnostics fail convergence and late-entry contacts exceed declared diagnostics. Slip still drops the payload. These failures remain in the committed measured summary and per-attempt receipts.

Scope: synthetic sensed-state mechanical coupons, without RGB perception, learned skill acquisition, full six-axis arm protection, physical sensor/material calibration, or hardware stopping qualification. Frozen placement/capsule runtime files are unchanged. CI runs cross-platform base tests plus native coupons, verifies evidence, and retains diagnostic artifacts.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant