Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 17 additions & 1 deletion .github/workflows/ci.yml
Original file line number Diff line number Diff line change
Expand Up @@ -62,9 +62,10 @@ jobs:
dvidia-skillspace-coverage --help
dvidia-episode-tape --help
dvidia-reaction --help
dvidia-motor-memory --help
dvidia-training-doctor --json
- name: Test footage, visual training and loopback studio
run: python -m unittest tests.test_footage_data tests.test_footage_train tests.test_footage_pipeline tests.test_footage_benchmark tests.test_training_studio tests.test_doctor tests.test_coverage tests.test_episode_tape tests.test_reaction tests.test_reaction_metrics tests.test_reaction_convergence tests.test_reaction_benchmark -v
run: python -m unittest tests.test_footage_data tests.test_footage_train tests.test_footage_pipeline tests.test_footage_benchmark tests.test_training_studio tests.test_doctor tests.test_coverage tests.test_episode_tape tests.test_reaction tests.test_reaction_metrics tests.test_reaction_convergence tests.test_reaction_benchmark tests.test_motor_memory tests.test_motor_metrics tests.test_motor_benchmark -v

arm:
name: Optional arm / Linux / Python 3.12
Expand Down Expand Up @@ -98,6 +99,21 @@ jobs:
name: reaction-mechanical-coupons
path: runs/reaction-ci
if-no-files-found: ignore
- name: Fit and verify motor memory and matched approach tasks
run: |
dvidia-motor-memory train --output runs/motor-memory-ci
dvidia-motor-memory inspect runs/motor-memory-ci
dvidia-motor-memory benchmark --memory runs/motor-memory-ci --output runs/motor-benchmark-ci --repeats 1
dvidia-motor-memory inspect runs/motor-benchmark-ci
- name: Preserve motor memory and task evidence
if: always()
uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
with:
name: motor-memory-approach-tasks
path: |
runs/motor-memory-ci
runs/motor-benchmark-ci
if-no-files-found: ignore

container:
name: Visual CLI container
Expand Down
10 changes: 10 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -98,6 +98,16 @@ The [reaction framework v0.1](docs/reaction-framework.md) adds a reusable monito

[The first measured run](benchmarks/reaction-v0.1-20261009/README.md) preserves 78 primary trials and eight timestep checks, including late-impact diagnostic failures and payload drops. Run `dvidia-reaction benchmark --output runs/reaction --repeats 3 --convergence`, then `dvidia-reaction inspect runs/reaction` to generate and verify your own offline report.

The [motor-memory framework v0.2](docs/motor-memory.md) fits a compact actuator model on separate native calibration episodes, then compares slow feedback, a fixed fast approach and predictive adaptive control on “approach, settle and retain.” It measures completion time against native error at the completion claim, overshoot, false/revoked completion, retained load and CPU cost. Commanded goals are exact; observations of the moving carriage can be delayed, noisy or missing. The movement profile and slowdown rules are authored. This is a CPU simulation experiment with learned dynamics, and does not train a task policy from human video or drive a physical robot.

```sh
.venv/bin/dvidia-motor-memory train --output runs/motor-memory
.venv/bin/dvidia-motor-memory benchmark --memory runs/motor-memory --output runs/motor-benchmark --repeats 3
.venv/bin/dvidia-motor-memory inspect runs/motor-benchmark
```

Each command uses a fresh output directory where applicable. The model and complete attempted matrix are frozen before evaluation; development and confirmation partitions remain separate. [The measured v0.2 run](benchmarks/motor-memory-v0.2-20261009/README.md) preserves the candidate, compact attempt receipts and limitations. Complete calibration transitions and scoring traces are retained in local runs and CI artifacts.

## Public pilot on Hugging Face

[Recorded demo](https://huggingface.co/spaces/Dvidia/dvidia-training) · [Synthetic datasets](https://huggingface.co/datasets/Dvidia/dvidia-training-examples) · [Pilot models and measurements](https://huggingface.co/Dvidia/dvidia-training-pilot)
Expand Down
85 changes: 85 additions & 0 deletions benchmarks/motor-memory-v0.2-20261009/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,85 @@
# Recorded motor-memory v0.2 experiment

This snapshot preserves a CPU-native **approach, settle and retain** experiment.
The learned component is a small actuator-dynamics fit. The movement profile,
speed governor, goals, completion and contact criteria are authored.

Four separate two-second calibration episodes produce **8,000 native 1 ms
transitions**. These are not 8,000 independent demonstrations or human videos.
Three stopping coupons supply an empirical braking assumption. Candidate,
protocol and runtime source hashes were frozen before the complete evaluation.
The six confirmation combinations were not executed during development; after
this run they are regression evidence for future revisions.

The measured run contains **324 attempts**: 12 cases × 3 controllers × 3 speed
tiers × 3 repeats. Every attempt ran the four-second horizon through native
physics. All 324 were numerically valid; runtime validity is separate from task
success and physical accuracy. The goal tolerance is 3 mm, speed tolerance is
5 mm/s and completion dwell is 50 ms.

## Confirmation outcomes

| Controller | All attempts | True task successes | False completion claims | Dropped loads | Median completion among verified successes | Median native error at that claim |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| Slow feedback | 54 | 36 | 0 | 9 | 2.005 s | 0.101 mm |
| Fast fixed | 54 | 36 | 0 | 9 | 1.730 s | 0.0109 mm |
| Adaptive predictive | 54 | 36 | 0 | 9 | 1.275 s | 0.181 mm |

The completion and error medians use the same 36 successful authored trials per
controller. Blackout and severe-slip cases contribute the remaining 18 attempts;
they remain incomplete. Slip drops nine loads per controller. An appropriate
latched hold is not credited as task success. Development has 54 attempts per
controller, with 45 successful completions and nine missing-input holds.

Adaptive control trades some native error at the claim for shorter completion in
the aggregate. It does not win every condition. Configurations combine an
authored speed profile, predictor and governor; component ablations are needed
before attributing gains to learned prediction alone. These exact-goal native
errors are not measurements of a physical robot's accuracy.

## CPU measurements and boundaries

The serial run simulated 1,296 seconds in 205.855 summed trial wall seconds:
**6.30 simulated seconds per wall second**. The complete benchmark loop including
JSONL writing took 207.158 seconds, excluding imports, calibration and final
report serialization. Median per-trial p95 controller-call cost on confirmation
was approximately 117 microseconds for adaptive control and 50 microseconds for
the feedback baselines. These measurements are not execution-time guarantees.

The host was Darwin 25.5.0, arm64, 10 logical CPUs, Python 3.12.9, NumPy 2.5.3 and
MuJoCo 3.15.0; no GPU was used. Process-lifetime peak RSS was 111,706,112 bytes,
including imports and the run. CPU model and total RAM were not identified by
this receipt, and no minimum hardware specification was established.

## Files and reproduction

- `memory.json`: frozen fitted candidate and calibration/braking identities.
- `summary.json`: compact protocol, source/host identities, split summaries,
frontier points and the complete run's evidence hashes.
- `attempts.jsonl`: all 324 scalar attempt receipts and trace lengths.

This folder is a compact snapshot, **not a complete inspectable run**. The full
training directory retains calibration rows; the full benchmark directory retains
every 10 ms scoring trace and controller-call timing sample. CI retains both as
the `motor-memory-approach-tasks` artifact. Reproduce from the current checkout:

```sh
.venv/bin/python -m pip install ".[arm]"
.venv/bin/dvidia-motor-memory train --output runs/my-memory
.venv/bin/dvidia-motor-memory inspect runs/my-memory
.venv/bin/dvidia-motor-memory benchmark --memory runs/my-memory --output runs/my-benchmark --repeats 3
.venv/bin/dvidia-motor-memory inspect runs/my-benchmark
```

Fresh output directories prevent replacing earlier evidence. A
`--development-only` benchmark excludes confirmation cases during development.
Cross-platform/runtime costs can differ; hashes establish consistency rather than
authenticated provenance. Repeated authored cases are correlated.

This coupon has one horizontal travel axis and pre-established parallel-jaw grip.
Commanded goals are exact, while robot-state observations can be delayed/noisy or
missing. It does not learn goal recognition, acquire a task from video, protect a
whole arm, perform recovery, or establish hardware qualification. The 0.28 m/s
tier is beyond the highest 0.24 m/s braking-calibration target; that is tested
extrapolation under an empirical envelope. Contact-force convergence remains
unresolved from v0.1. See [the framework contract](../../docs/motor-memory.md).
Loading
Loading