Repository navigation
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The reaction monitor can stop motion, but it does not demonstrate fast, precise task execution. This adds motor-memory v0.2: separate native actuator calibration followed by an offline “approach, settle and retain” benchmark comparing slow feedback, fixed fast approach and adaptive predictive control at three matched speed tiers.
The learned component is a compact actuator-dynamics fit on 8,000 transitions from four separate two-second calibration episodes. Runtime prediction consumes timestamped sensed state and acknowledged applied commands. The motion profile, governor and outcome rules are authored. Commands retain their application delay; independent reaction holds latch and native physics continues to the same four-second horizon in every trial.
Evidence separates development and confirmation, records every planned attempt, checks native truth at observed completion claims, and reports error/overshoot, false or revoked claims, later rest loss, retained/dropped loads, contact proxies, sensing faults, controller latency and CPU throughput. Inspection refits calibration, verifies identities and the canonical attempted matrix, and recomputes outcomes and aggregates from complete receipts. The offline HTML plots completion time against native error at the claim and links complete traces. CI verifies both training and benchmark receipts and retains artifacts.
Validation: 287 local tests passed; installed CLI/dependencies and complete evidence inspection passed. The frozen recorded run has 324 native attempts with no numerical/runtime errors. Each controller has 54 confirmation attempts: 36 successful tasks, 18 blackout/slip failures, no false completion claims and nine dropped loads. The same successful-trial cohort has median completion 1.275 s for adaptive control versus 1.730 s fixed and 2.005 s slow; adaptive trades some native claim error for speed. Serial throughput was 6.30 simulated seconds per wall second on the recorded CPU without a GPU. Raw local traces and compact committed receipts retain all failures.
Scope: one-axis carriage and pre-established parallel-jaw grip, exact commanded goals and synthetic delayed/noisy robot-state sensing. This is fitted dynamics plus authored control, without human-video task-policy acquisition, camera goal localization, full-arm geometry, recovery or hardware qualification. The 0.28 m/s tier extrapolates beyond the 0.24 m/s braking calibration, and prior impact-force convergence remains unresolved. Complete configurations need component ablations before attributing speed gains to learning alone.
This PR builds on #4 and targets its v0.1 branch so the review shows only v0.2 additions. Existing reaction/placement source and v0.1 evidence are unchanged.