Project 13 · requirements version 1
A simulated robot that learns a task
Collect demonstrations, train a bounded policy, compare single actions, chunks and generative routes, then freeze a candidate before opening separate evaluation combinations. This project saves local evidence in progress backups. It does not authenticate authorship or certify hardware safety.
Storage: loading. Report: draft. Independent human artifact review: pending.
1. Collect and compare
Record and save a learner trajectory in the workbench. Train the motor policy, save single-action and chunk captures with identical weights/settings, then train and save a generative-route capture. Failed outcomes remain useful evidence. The lab is temporary until you save its configuration here.
Planar robot learning · CPU teaching experiment
A puck is translated by a pusher kept in ideal permanent contact. Commands are velocities in metres/second, held for 0.1 s. The plant moves by 0.1 × (1 − drag) × command. This model omits impact, slipping contact, rotation and inertia. The square workspace is fixed at ±1.2 m; object radius counts toward wall collisions.
Demonstration recorder and receipt audit
Position (-0.6500, 0.0000) m; 0/30 commands; 0/4 saved learner trajectories. Recording settings freeze at the first command (drag 0.2); later controls affect a new trajectory. Recorder observations are exact; noise and delay apply only to supplied collection and policy evaluation.
| Time before | Observation (m) | Command (m/s) | Next position (m) |
|---|
Supplied episodes: eight training × twenty commands, eight separate development starts. Learner trajectories retain their own settings and origin. Fitting accepts only training split labels; imported histories are replayable records, not authenticated provenance.
Learn actions from rendered positions
The 6–8–6 tanh network has 110 trainable parameters. Inputs are current and goal x/y coordinates plus one-hot tokens “left”/“right”. Six outputs predict three consecutive two-coordinate commands. The instruction vocabulary has exactly two known tokens; coordinates are supplied from the renderer, not inferred from pixels. The teacher is a programmed proportional controller.
Radius changes wall checks, not teacher actions; this training collection has no wall contacts. Randomizing a parameter without relevant examples does not teach its effects.
Updates 0/400; mean training half-squared command error 0.1928 (m/s)² over 160 rows × six outputs. All gradients use the same preceding weights. Low supervised loss does not establish closed-loop success.
Solid: learned policy; dashed: programmed teacher. Equal axis scales; exact numbers below. Both receive the same seeded sensor offsets and imposed disturbance, but their feedback trajectories differ. The illustrated puck is the selected receipt, not a new simulation.
Observation from interval 0: (-0.5000, -0.2250). Command (-0.0797, -0.0209) m/s. Actual next (-0.5064, -0.2267) m. Goal distance 1.2609 m.
| Outcome | Learned | Teacher |
|---|---|---|
| Status | time-limit | success |
| Final distance (m) | 1.4422 | 0.0021 |
| Collision | false | false |
| Observations | 30 | 30 |
| Policy calls | 30 | 0 learned calls |
Success requires all thirty intervals and the last three within 0.12 m, without collision. Time limits remain failures in the denominator. Simulated time is 3.0 s, not CPU latency; browser rendering, fitting and communication overhead are unmeasured.
Train and inspect conditional action denoising
Training contains two route modes, two condition offsets and twelve Gaussian examples per mode/condition/noise level (192 pairs). The scalar route coefficient becomes thirteen planar waypoints; differences are an action sequence. No images or full robot dynamics are learned by this denoiser.
0/400 updates; training half-squared error 0.2112.
Lateral coefficient 0.0000 m. Every segment clears the radius-0.25 m obstacle: no. Dashed route is the zero-coefficient mean of symmetric clean demonstrations and intersects the obstacle. Sampling can also fail: verification, not model loss, decides clearance.
| α | Noisy input | Predicted clean | Next |
|---|---|---|---|
| 0.15 | -1.0000 | 0.0000 | -0.8402 |
| 0.4 | -0.8402 | 0.0000 | -0.5941 |
| 0.7 | -0.5941 | 0.0000 | -0.2425 |
| 0.95 | -0.2425 | 0.0000 | 0.0000 |
All work is bounded direct CPU work; no background animation.
2. Inspect saved configurations
0/6 captures. Scores are recomputed, never imported as authoritative results.
3. Evaluate the frozen candidate
The four evaluation combinations are separate from fitting. Public source code makes these inspectable; this is self-administered evidence. Once viewed, later comparisons are practice. Exposure survives resets/imports. Freezing a new candidate does not erase earlier exposure.
Freeze a saved single-action or chunk configuration first.
4. Explain and submit
Provenance: who/what supplied demonstrations, timestamps, units and split. Comparison: matched baselines and denominators. Disturbance: recovery and failure receipts. Embodiment: planar/3D/hardware differences. Decision: justified acceptance, unresolved risks and additional evidence.
Code never executes here. Implementation requires runnable tests and standard-simulator evidence; report presence checks do not assess their quality.
- Save at least one learner-recorded demonstration.
- Save single-step and chunk results from the same trained policy and settings.
- Save a trained generative route comparison.
- Freeze a policy and evaluate the separate scenario combinations.
- Write the provenance report.
- Write the comparison report.
- Write the disturbance report.
- Write the embodiment report.
- Write the decision report.
5. Separate objective transfer cases
These new scalar/design scenarios use independent oracles. Four criteria each; three attempts maximum. First attempts remain distinct from practice retries. They assess bounded reasoning, not code quality or hardware proficiency.