No approved smooth teacher yet. The latest stronger-penalty trial succeeds in46.1% of episodes. Its median acceleration measure is71% below the retained reference, but the fast initial movement and inaccurate holds remain. The retained teacher is preserved; the other23 teachers are paused.
| Policy | Success | Peak speed¹ (rad/s) | Acceleration¹ median / p95 (rad/s²) | Review |
|---|---|---|---|---|
| Retained reference | 118/128 +117/128 | 8.07 | 56.95 /58.92 | Before in every pair |
| S13 · Stronger target-distance penalty | 59/128 (46.1%) | 7.50 | 16.35 / 33.71 | Paired videos · W&B |
| S12 · Lower target-distance penalty | 3/128 (2.3%) | 8.20 | 15.97 / 31.56 | Paired videos · W&B |
| S11 · Separate gradient clipping | 11/128 (8.6%) | 8.37 | 16.30 / 27.02 | Paired videos · W&B |
These candidates fail the success and motion requirements. Their median episode duration is20 seconds. Most still grasp and lift the cube, but often hold it outside the required goal region.
Videos are1× real time: 50 frames/s, up to20 seconds. Each recorded20ms state is rendered once. Three fixed episodes use identical initial conditions and camera views; failures are included.
¹ All128 first episodes are measured, including failures. Peak speed is the95th percentile of each episode's sampled maximum joint speed. Acceleration is an RMS proxy from20ms velocity samples; the table shows its median and95th percentile.
Training now: S14/S15 test continuous goal reward weights50/500 while retaining the motion penalties. Both passed checks with a smaller exploration increase and are training on GPUs1/2. Final success evaluation and real-time videos are pending; neither is approved.
Earlier partial progress: S7,103/128 with less median jitter · S8 · S9,37/128 with acceleration p95 reduced57% · S10,1/128. None is an approved replacement.
Earlier penalty trials: 10× · 100× · Early-collision failure · S1 · S2 · S3.