| --- |
| license: mit |
| pretty_name: Microduck skill tree (sim-only training log) |
| tags: |
| - robotics |
| - microduck |
| - reinforcement-learning |
| - mujoco |
| - mjlab |
| - simulation |
| size_categories: |
| - n<1K |
| --- |
| |
| # Microduck skill tree, the training log |
|
|
| Simulation only. Nothing here has run on a real robot yet; the Microduck this is for is still on order. Every policy, curve and clip in this repo comes from mjlab / MuJoCo Warp on one RTX 5090, trained with the tasks in [pollen-robotics/microduck_rl](https://github.com/pollen-robotics/microduck_rl) at commit `53b8971`. |
|
|
| This is the build log. One folder per level of the skill tree, publishable or not, partial or not. The installable policies live in their own model repos (one per policy, in the shape `robotctl` installs); this dataset holds everything around them: the reward curve of every run, the exact queue item that ran it, the proof-take logs the verdict was read from, the clips, and the manifests. |
|
|
| Sibling dataset: [witcheer/microduck-hermes-runs](https://huggingface.co/datasets/witcheer/microduck-hermes-runs) logs every run where Hermes Agent drives the simulated duck from one typed sentence, with the agent's full transcript, the simulator's ground truth and a replay video. Both datasets and every policy repo are grouped in the [Microduck skill tree collection](https://huggingface.co/collections/witcheer/microduck-skill-tree-sim-only-6ab195da0c0b293bdccc6d34). |
|
|
| ## The ladder |
|
|
| | level | skill | task id | iterations | wall time | verdict in sim | policy repo | |
| |---|---|---|---|---|---|---| |
| | 1 | walk | `Mjlab-Velocity-Flat-MicroDuck` | 4,000 | 58 min (0.88 s/it) | 0.15 m/s held for 12 s on a 0.30 m/s command, trunk 119 to 121 mm | [microduck-walk-flat](https://huggingface.co/witcheer/microduck-walk-flat) `@v1` | |
| | 2 | get up | `Mjlab-StandUp-Flat-MicroDuck` | 25,000 (15k + 10k extension) | 407 min (0.98 s/it) | stands from sitting and from face down and holds; from flat on its back it freezes half-rolled, 0 of 2 takes | [microduck-standup](https://huggingface.co/witcheer/microduck-standup) `@v1` | |
| | 3 | sit and stand on command | `Mjlab-SitStand-Flat-MicroDuck` | 15,000 | 272 min (1.08 to 1.10 s/it) | stand at 116 mm, sit at 59 mm, switches in under 1 s in the takes | none: posture-flag policies are daemon-driven and not publishable through `uv run publish` | |
| | 4 | walk (rough-terrain task, random rough patches; see the model card) | `Mjlab-Velocity-Rough-MicroDuck` | 8,000 | 324 min (2.44 s/it) | upright for a 12 s take on flat ground; on the hardest terrain row (takes recorded 2026-10-11) 3 of 3 without gear play, 4 of 9 with it | [microduck-walk-rough](https://huggingface.co/witcheer/microduck-walk-rough) `@v1` | |
| | 5 | walk and recover in one policy | `Mjlab-VelStand-Flat-MicroDuck` | 20,000 | 326 min (0.98 s/it) | walks and holds 20 s from standing, catches a forward lean; does not get up from flat on its back, 0 of 3 | [microduck-walk-recover](https://huggingface.co/witcheer/microduck-walk-recover) `@v1` | |
| | 6 | forward roll | `Mjlab-Roulade-Flat-MicroDuck` | 10,000 | 173 min (1.04 s/it) | 3 of 3 takes roll and finish standing: on its back at 0.8 s, upright at 1.0 s, 115 mm by 1.2 s | [microduck-roulade](https://huggingface.co/witcheer/microduck-roulade) | |
| | 7 | right-foot ball kick | `Mjlab-BallKick-Flat-MicroDuck` | 10,000 | 161 min (0.97 s/it) | 3 of 3 takes kick and stay standing: the ball travels 1.03 to 1.13 m in the first second, trunk 100 to 123 mm over the 4.1 s logged, no resets; the policy never sees the ball (by design), so the ball has to sit at the right toe | [microduck-kick-right](https://huggingface.co/witcheer/microduck-kick-right) | |
| | 1b | walk, with ±1° of gear play in every servo | `Mjlab-Velocity-Flat-Backlash-MicroDuck` | 3,000 on top of level 1's 4,000 | 57 min (1.16 s/it) | with gear play: on its feet for 20 s in 4 of 5 takes, level 1 in 2 of 5 (level 1 in its own sim without gear play: 1 of 3); the training fall rate is no better (fell_over 0.496 against 0.453), so a small-sample hint only | [microduck-walk-flat](https://huggingface.co/witcheer/microduck-walk-flat) `@v2` (default) | |
| | 5b | walk and recover, with ±1° of gear play in every servo | `Mjlab-VelStand-Flat-Backlash-MicroDuck` | 3,000 on top of level 5's 20,000 | 62 min (1.25 s/it) | with gear play: 6 of 7 standing takes hold 20 s, level 5 also 6 of 7; neither gets up from its belly or back, 0 of 3 each | [microduck-walk-recover](https://huggingface.co/witcheer/microduck-walk-recover) `@v2` (default) | |
| | 2b | get up, with ±1° of gear play in every servo | `Mjlab-StandUp-Flat-Backlash-MicroDuck` | 3,000 on top of level 2's 25,000 | 63 min (1.28 s/it) | with gear play, the same as level 2 with the same gear play in every pose: face down 3 of 3 (standing, leaning forward 31 to 37°), sitting 3 of 3, standing 3 of 3, on its back 0 of 3 (half-rolled freeze) | [microduck-standup](https://huggingface.co/witcheer/microduck-standup) `@v2` (default) | |
| | 3b | sit and stand on command, with ±1° of gear play in every servo | `Mjlab-SitStand-Flat-Backlash-MicroDuck` | 3,000 on top of level 3's 15,000 | 71 min (1.43 s/it) | with gear play: no fall in 4 of 5 takes of 25 s, level 3 with the same gear play in 3 of 5 (small sample); sits at about 60 mm, stands at about 115 mm, 1 to 4 switches per take | none: posture-flag policies are daemon-driven and not publishable through `uv run publish` | |
| | 5c | walk and recover on rough ground | `Mjlab-VelStand-Rough-MicroDuck` | 3,000 on top of level 5's 20,000 | 183 min (about 3.7 s/it) | from standing on the hardest terrain row: on its feet 20 s in 7 of 9 takes, level 5 in 6 of 9 (stairs 2 of 3 against 0 of 3); from its belly or back 0 of 9, level 5 also 0 of 9 | [microduck-walk-recover-rough](https://huggingface.co/witcheer/microduck-walk-recover-rough) | |
| | 4b | walk on rough ground, with ±1° of gear play in every servo | `Mjlab-Velocity-Rough-Backlash-MicroDuck` | 3,000 on top of level 4's 8,000 | 147 min (about 2.9 s/it) | with gear play, hardest terrain row: on its feet 20 s in 8 of 9 takes, level 4 in 4 of 9 (stairs 2 of 3 against 1 of 3, block grid 3 of 3 against 1 of 3, slope 3 of 3 against 2 of 3); without gear play level 4 3 of 3, level 4b 2 of 3. small sample, and two things changed (gear play and the ordered terrain rows) | [microduck-walk-rough](https://huggingface.co/witcheer/microduck-walk-rough) `@v2` (default) | |
| | 11 | walk, and get up from flat on its back | `Mjlab-VelStand-Flat-MicroDuck` on the maker's public `soft_carpet` branch (commit `695f3b7`, its distillation recipe: imitate a get-up policy on fallen frames, keep the walk on upright ones) | 6,000, then 3,000 more on deeper carpet | 141 min (1.41 s/it), then 73 min (1.48 s/it) | 9 forced-spawn takes of 12 s, with the task's own shoves knocking it over every 1 to 2 s: 34 times on the ground, back up from 30, median 0.8 s; level 5 in the same sim 0 of 3. On 10 mm carpet with the maker's motor model it goes down by 4 s and stays down, before and after the deeper-carpet 3,000 | none: the gear-play version, level 11c, is the published one | |
| | 11c | level 11, with ±1° of gear play in every servo | `Mjlab-VelStand-Flat-Backlash-MicroDuck` (`soft_carpet`) | 3,000 on top of level 11's 6,000 | 83 min (1.68 s/it) | with gear play, 7 forced-spawn takes of 12 s: 35 times on the ground, back up from 31, upright 59% of the time; on 10 mm carpet with the motor model it fell at 3 s and got back up once (level 11 stayed down): less stuck, not fixed | [microduck-walk-getup](https://huggingface.co/witcheer/microduck-walk-getup) `@v1` (the get-up is Pollen's `alpha_stand` expert, distilled with their branch recipe) | |
| | 11d | walk and get up on rough ground, with ±1° of gear play | `Mjlab-VelStand-Rough-Backlash-MicroDuck` (`soft_carpet`, the task the maker's own config calls its production recipe) | 3,000 on top of level 11c's 9,000 | 209 min (about 4.2 s/it) | not better than 11c anywhere. hardest row from standing (9 takes each): 50 times down, 43 rises, upright 42% of the time, against 11c's 55 / 48 / 51%; row 6 from belly or back (6 takes each): 32 / 27 / 37% against 39 / 36 / 46%; flat with gear play 26 / 21 / 45% against 35 / 31 / 59%. Mean reward flat around -17 for the whole run | none: no gain over level 11c | |
| | 1c | walk with up to +60 g on its back, with ±1° of gear play | `Mjlab-Velocity-Flat-Backlash-Payload-MicroDuck` (our fork: trunk mass randomised over 0.95 to 1.30x instead of the maker's 0.95 to 1.05x, one parameter) | 3,000 on top of level 1b's 7,000 | 57 min (1.16 s/it) | not better than level 1b anywhere, 20 s takes: trunk at 1.30x (+60 g) with shoves 3 of 5 against 3 of 5, 1.30x without shoves 3 of 3 against 3 of 3, normal mass with shoves 3 of 5 against 4 of 5. Level 1b already carries the extra 60 g in sim; small samples | none: no gain over level 1b | |
| | 1d | walk on slippery floors, with ±1° of gear play | `Mjlab-Velocity-Flat-Backlash-Slip-MicroDuck` (our fork: foot friction randomised over 0.2 to 1.3 instead of the maker's 0.7 to 1.3, one parameter) | 3,000 on top of level 1b's 7,000 | 58 min (1.15 s/it) | 20 s takes against level 1b: foot friction 0.2 with a shove every 0.5 to 1 s, on its feet in 8 of 15 against 1 of 15 (one-sided Fisher exact p = 0.0071); 0.2 without shoves 12 of 13 against 8 of 13 (p = 0.080, a lean); normal friction with shoves 9 of 15 against 8 of 15, no difference | [microduck-walk-slip](https://huggingface.co/witcheer/microduck-walk-slip) `@v1` (walk-flat stays the default walk: they tie at normal grip) | |
| | 1e | walk over a single step, with ±1° of gear play | `Mjlab-Velocity-Flat-Backlash-Step-MicroDuck` (our fork: the maker's flat plane becomes a floor with one 1.5 m platform per 4 m tile, its top 10 to 25 mm above the floor, one change; the policy still has no terrain scan) | 3,000 on top of level 1b's 7,000 | 64 min (1.28 s/it) | 20 s takes against level 1b, walking straight ahead at 0.3 m/s off a 25 mm step: without shoves on its feet in 11 of 15 against at most 5 of 15 (one-sided Fisher exact p = 0.033); with a shove every 0.5 to 1 s 5 of 15 against 2 of 15 (p = 0.19); both pooled 16 of 30 against at most 7 of 30 (p = 0.016). Same layout at 1 mm with shoves 10 of 15 against at most 7 of 15 (two-sided p = 0.46; 'at most': walk-flat v2 takes cut short upright at the tile-grid edge count as held), so part of the gain may be general sturdiness. No take reached the next platform, so the step up is untested | [microduck-walk-step](https://huggingface.co/witcheer/microduck-walk-step) `@v1` (walk-flat stays the default walk) | |
| | 1g | walk on gentle hills, with ±1° of gear play | `Mjlab-Velocity-Flat-Backlash-Slope-MicroDuck` (our fork: the maker's flat plane becomes a grid of 4 m tiles with one pyramid hill each, slope 0.05 to 0.15 rise per run, one change; no terrain scan) | 3,000 on top of level 1b's 7,000 | about 128 min (2.55 s/it, over two sittings) | 20 s takes against level 1b, walking straight ahead at 0.3 m/s off a 0.15 hilltop: both hill conditions pooled (the read picked in advance) 23 of 30 against 12 of 30 (one-sided Fisher exact p = 0.0041); without shoves 14 of 15 against 8 of 15, with the maker's play shoves 9 of 15 against 4 of 15 (p = 0.070, a hint alone); flat 1 mm control 10 of 15 each, no loss. Only the 0.15 hill was tested | [microduck-walk-slope](https://huggingface.co/witcheer/microduck-walk-slope) `@v1` (walk-flat stays the default walk) | |
| | 1h | one walk for a step, a slippery floor and harder shoves at once, with ±1° of gear play | `Mjlab-Velocity-Flat-Backlash-Mix-MicroDuck` (our fork: 1e's step floor, 1d's foot friction 0.2 to 1.3 and the walk-push fork's doubled training shove together; it imports their task files) | 3,000 on top of level 1e's 10,000 | 64 min (1.29 s/it) | 20 s takes walking straight ahead at 0.3 m/s. All three at once against level 1b (picked in advance): 9 of 15 against 0 of 15 (one-sided p = 0.0003). No measurable loss against each single-condition policy on its own condition: step 10 of 15 against walk-step's 13 of 15 (two-sided p = 0.39), doubled shoves 10 of 15 against walk-push's 9 of 15 (p = 1.0), slippery floor with shoves 11 of 15 against walk-slip's 3 of 15 (p = 0.0092), a low draw for walk-slip (6 to 8 of 15 in three other sets), so no loss, not a win | [microduck-walk-mix](https://huggingface.co/witcheer/microduck-walk-mix) `@v1` (walk-flat stays the default walk) | |
| | 1i | level 1h plus gentle hills in one walk, with ±1° of gear play | `Mjlab-Velocity-Flat-Backlash-MixSlope-MicroDuck` (our fork: each 4 m tile is either 1e's single 10 to 25 mm step or 1g's pyramid hill of slope 0.05 to 0.15, with 1h's foot friction 0.2 to 1.3 and doubled shoves; it imports their task files) | 3,000 on top of level 1h's 13,000 | 120 min (2.41 s/it) | no measurable gain over level 1h. 20 s takes walking straight ahead at 0.3 m/s. On the condition picked in advance (0.15 hill + 25 mm step + friction 0.2 + doubled shoves) 5 of 15 against walk-mix's 3 of 15 (one-sided Fisher exact p = 0.34). No loss on the other two: 0.15 hill without shoves 15 of 15 against walk-slope's 15 of 15; walk-mix's own floor 7 of 15 against walk-mix's 5 of 15 (two-sided p = 0.71), inside the noise | [microduck-walk-mixslope](https://huggingface.co/witcheer/microduck-walk-mixslope) `@v1` (walk-flat stays the default walk, walk-mix the mixed-floor walk) | |
| | 8 | beak to the ground without touching it | `Mjlab-GroundPick-Flat-MicroDuck` | 20,000 planned | queued | | none: phase-driven, daemon-side | |
|
|
| All budgets are the maker's `max_iterations` from the task config unless stated. Levels 1b, 2b, 3b and 5b are the maker's backlash variants of levels 1, 2, 3 and 5: a free hinge of ±1° in series with each of the 14 servo joints, with the joint encoder reading through the play as a real servo does. Each started from its level's final checkpoint, and each publishable one is v2 of that level's policy repo (v1 stays installable with `@v1`). Level 4b is the same for level 4 and is v2 of walk-rough. Levels 11, 11c and 11d were trained on the maker's public `soft_carpet` branch of microduck_rl (commit `695f3b7`), not on `53b8971`; 11 and 11c on flat ground, 11d on the ordered terrain ladder with the same local patches as 5c. Level 5c starts from level 5's final checkpoint and trains on the rough version of the same task, on a terrain laid out as an ordered difficulty ladder and with five local patches to the maker's code (in its `patches/` folder; the model card says what each does). Levels 1c and 1d are our own conditions, not the maker's: each is a separate task file in microduck_rl (a copy is in the level's `patches/` folder) that changes one parameter of `Mjlab-Velocity-Flat-Backlash-MicroDuck` and is registered by one import line, and each started from level 1b's final checkpoint (walk-flat v2). 4096 parallel environments, 24 steps per iteration, PPO from rsl_rl, tensorboard logging. |
| |
| ## What is in a level folder |
| |
| - `reward-curve.csv`: every scalar rsl_rl logged per iteration (`Train/mean_reward`, `Train/mean_episode_length`, `Perf/total_fps`, the `Episode_Termination/*` terms), pulled from the tfevents of every run of that level and concatenated by iteration. |
| - `reward-curve.png`: mean reward per iteration, raw and a 50-iteration moving average. |
| - `queue-item.sh`: the exact header that trained it. The shared body it sources is `tools/duck-train-body.sh` (resumable, checkpoint every 100 iterations, exports with the maker's `scripts/export.py`). |
| - `manifest.json`: the schema-2 manifest `uv run publish` wrote for the policy repo, when there is one. |
| - `proof-takes/*.log`: the recorder's per-step trunk height and body-frame gravity for each take the verdict was read from (levels 5, 6, 7, 1b, 2b, 3b, 5b, 5c, 4b, 11, 11d, 1c, 1d, 1e, 1g, 1h and 1i so far; level 7 adds the ball position on every line). Level 5b keeps one subfolder per recording batch: forced spawns standing, face down and face up for both policies (`vsbl-matrix-*`), a random-spawn batch (`vsbl-2026-09-26`), trunk-tracking takes (`vsbl-track-*`), and level 5 in its original sim without gear play (`vsbl-ref-*`). In the 1b folder `seed-*` is level 1, `twin-*` is 1b, and `ref-seed-*` is level 1 in its original sim; 2b and 3b use `seed-*` / `twin-*` / `ref-*` the same way. In 5c, `proof-vsr3-2026-10-02` has `v3-*` (5c) and `seed-*` (level 5) on forced terrain rows, and `vsrough3-terrain-2026-10-01` compares the 22,900-iteration checkpoint (`vsr3-*`) against a rough retrain on the stock random grid (`vsr2-*`). In 4b, `proof-y-a-2026-10-11` has `y-*` (4b) and `v1-*` (level 4) on forced rows 6 and 9 with gear play, and `proof-y-b-2026-10-11` has both in the plain rough sim (`ref-*`). In 11d, `x-*` is 11d and `w-*` is 11c: `proof-x-a` standing on row 9, `proof-x-b` from belly and back on row 6, `proof-x-c` flat with gear play. Level 11's `verdict-carpet-*` folders are the flat forced-spawn takes, and `carpet-gap-*` (also in 11c) are the maker's `infer_policy.py` on soft carpet, one full log per case. Level 11c's flat forced-spawn takes, the ones its model card quotes, are the `w-*` files in 11d's `proof-x-c-2026-10-11`. In 1c and 1d, `z-*` (1c) and `z2-*` (1d) are the new policy and `v2-*` is level 1b, one subfolder per condition: in 1c `a` = trunk mass 1.30x with play shoves, `b` = 1.30x without shoves, `c` = normal mass with shoves; in 1d `a` = foot friction 0.2 with shoves, `b` = 0.2 without shoves, `c` = 1.0 with shoves, and each `*x` folder is the extension of that condition (the take numbers continue from the first set). In 1e, `z3-*` is the new policy and `v2-*` is level 1b, every take walking straight ahead at 0.3 m/s from a standing start on a platform: `a` = a 25 mm step with play shoves, `b` = the 25 mm step without shoves, `c` = the same layout at 1 mm with shoves (a zero-height box does not compile, so 1 mm is the no-step control). In 1g, `z5-*` is the new policy and `v2-*` is level 1b, every take walking straight ahead at 0.3 m/s from a hilltop: `a` = 0.15 hill with the maker's play shoves, `b` = 0.15 hill without shoves, `c` = the same layout with 1 mm hills (control); `proof-takes/score.txt` is the scorer's output. Level 1b's videos for `a` were lost to a temp-folder clash between two batches running in parallel; its take logs are complete. In 1h, `z6-*` is the new policy against a rival per condition: `a` = 25 mm step + friction 0.2 + doubled shoves every 3 to 6 s against level 1b (`v2-*`), `b` = 25 mm step, friction 1.0, no shoves against walk-step (`z3-*`), `c` = 1 mm, friction 0.2, the maker's shove against walk-slip (`z2-*`), `d` = 1 mm, friction 1.0, doubled shoves every 3 to 6 s against walk-push (`z4-*`); `diag-z2c-*` are the two walk-slip re-runs behind the `c` caveat (`dg1-*` its own plane with the fixed command, `dg2-*` the 1h floor with random commands). In 1i, `z7*-*` is the new policy against a rival per condition, every take walking straight ahead at 0.3 m/s: `a` = 0.15 hill + 25 mm step + friction 0.2 + doubled shoves every 3 to 6 s in the 1i task against walk-mix (`z6a-*`), `b` = 0.15 hill without shoves in the 1g task against walk-slope (`z5b-*`), `c` = 1h's own condition (25 mm step, friction 0.2, doubled shoves) in the 1h task against walk-mix (`z6c-*`); `proof-takes/score.txt` is the scorer's output. On a hill-top start `trunk_z` goes below zero once the duck walks down off the hill (it is measured from the spawn point): that is the descent, not a fall. Verdicts are read from these numbers, not from the video. |
| - `patches/*.py` (levels 5c, 1c, 1d, 1e, 1g, 1h and 1i): the local changes to the maker's code the level trained with; for 1c, 1d and 1e that is the fork task file. 1g carries its fork task file and `ground_height_patch.py` (height terms measured from the ground under the feet, not the tile origin, which is what lets them see a slope); 1h carries its fork file, the three single-condition task files it imports, and the same ground patch. 1i carries its fork file, the four task files it imports (step, slope, slip, push) and the same ground patch. |
| - `clip-*.mp4`: iteration 0 against the trained policy, same camera, same spawn, for levels 1, 5 and 6; both gear-play v2 walks side by side for level 1b, level 5 against 5b (both running with gear play) for level 5b, level 4b against level 4 on the hardest block-grid row with gear play (level 4 tips over at 7.1 s; the last frame is held 1.5 s), level 1d against level 1b at foot friction 0.2 with shoves for level 1d (a red box marks each fall, where the sim restarts the duck), level 1e against level 1b walking off a 25 mm step without shoves for level 1e (red box where level 1b falls and the sim restarts it, green box where level 1e walks down the step), the trained policy alone for levels 3 and 7 (level 7 holds the opening frame for about a second before the take plays, and the camera follows the ball). Level 1i's clip is the trained policy alone on a separate 40 s take (0.15 hill, 25 mm step, normal grip, no shoves; none of the scored 20 s takes crossed both tile kinds): it steps down off the step and later walks onto a hill, played at 2x, with no reset inside the window. Level 2 has no clip because the first one I made was wrong (an environment reset at the 6 s episode boundary read as a stand-up) and a correct one is not recorded yet. |
|
|
| ## Tools |
|
|
| - `tools/headless_play.py`: records any checkpoint of any mjlab task over ssh with no display. Both of mjlab's play viewers fail headless (viser asserts on a degenerate command slider, the native one wants X11); this wrapper swaps them for a plain policy loop, lets play's own recorder write the mp4, forces the spawn state, overrides the episode length, prints every reset and samples trunk height and body-frame gravity at a chosen cadence. Copy it into the microduck_rl project dir and run with `uv run python`. |
| - `tools/velbl-takes.sh`, `tools/vsbl-*.sh`: the recording batches behind levels 1b and 5b. `tools/proof-*.sh` and `tools/proof-verdicts.py`: the batches and the scorer behind 2b, 3b and 5c; `tools/vsrough3-terrain-matrix.sh` and `tools/vsrough-falls.py`: the earlier 5c terrain comparison. `tools/proof-common-wt.sh`, `tools/proof-y-*.sh`, `tools/proof-x-*.sh`, `tools/take-score-2026-10-11.py` and `tools/recov-gz-2026-10-11.py`: the batches and scorers behind 4b and 11d (the second counts falls and rises from body-frame gravity only, because on rough ground the trunk height is measured from the patch origin); `tools/verdict-carpet.sh`: level 11's flat takes. `tools/proof-z-*.sh` and `tools/proof-z2-*.sh`: the batches behind 1c and 1d (scored with `take-score-2026-10-11.py`); `tools/mass-play-check-2026-10-12.sh` and `tools/fric-play-check-2026-10-13.sh` prove the trunk-mass and foot-friction overrides reach the play env before any take counts. `tools/proof-z3-*.sh` and `tools/z3-score-2026-10-14.py`: the batches and the scorer behind 1e (the scorer also places each fall by its distance from the platform centre, 0.65 to 1.15 m being the edge band); `tools/step-play-check-2026-10-14.sh` checks the step-height override in the play env. `tools/proof-z5-*.sh`, `tools/z5-score-2026-10-16.py`, `tools/slope-play-check-2026-10-14.sh` and `tools/slope-calib-2026-10-14.sh`: the batches, scorer, slope-override check and control takes behind 1g. `tools/proof-z6-*.sh`, `tools/z6-score-2026-10-17.py`, `tools/upright-reset-check-2026-10-22.py`, `tools/worstcase-2026-10-22.py` and `tools/diag-z2c-*.sh`: the batches, scorer, upright-reset checks and walk-slip re-runs behind 1h. `tools/proof-z7-*.sh` and `tools/z7-score-2026-10-22.py`: the batches and scorer behind 1i; `tools/clip-z7-fresh.sh` and `tools/cut-mixslope-x.sh` record and cut its clip. |
| - `tools/duck-train-body.sh` and `tools/duck-lane.sh`: the queue item body and the weekday duck/bench alternation. Written for one specific box (paths under `/home/witcheer`); read them for the shape, not to run as-is. |
| |
| ## Things learnt on the way, in order |
| |
| - A queue item that adds a rendering flag needs a real probe of the exact argv, not a help parse: `--video` swallowed the next token (tyro wants `--video True`), and the offscreen renderer needs `MUJOCO_GL=egl` set by the item itself. Two training blocks lost. |
| - A resumed rsl_rl run counts the loaded iteration as its first, so `--agent.max-iterations TARGET-IT` stops one short of TARGET forever. Use `TARGET-IT+1`. |
| - A recording longer than the task's episode contains a reset that teleports the duck to a fresh spawn and reads as the skill happening. Know `episode_length_s` before recording; override it for the clip; print every reset. |
| - Timed terminations do the same thing periodically: VelStand resets a duck that has been down for 8 s, so every "get-up" landing exactly 8 s after a fall was a reset. Read `terminations[` in the task config before reading any recovery off the numbers. |
| - Read verdicts from trunk height and body-frame gravity, not from frames. Vision reads of small contact sheets were wrong three times in one afternoon. |
| - A one-second skill needs 0.1 s sampling; at 0.5 s the roll showed as upright, upright, nothing in between. |
| - The roll task deletes the fall termination (a roll starts with a fall) and spawns half its episodes mid-roll with the chin tucked, so the hard second half is practised from iteration 1. Its reward was flat from iteration 2,000 of 10,000. |
| - The kick task gives the policy no ball observation at all, because the real robot has no ball sensor; only the critic sees the ball. The ball is placed at the right toe at every reset and the policy kicks on timing. Its reward reached 92.9 by iteration 1,000 and ended at 90.1 at 10,000. 14 of its 10,001 logged iterations have a negative mean reward (the lowest -4,686.8 at iteration 4,361) while the mean episode length stays at about 246 to 250 steps, so these are not mass falls; the cause is not investigated yet. The chart clips the 8 deepest and says so; every raw value is in the csv. |
| - A moving target needs a camera that follows it. A trunk-tracking camera lost the ball by 1.5 s, a fixed camera lost it out of a different edge each take (the spawn heading is random), and a wide shot shrank the duck to a speck. Tracking the ball at 0.9 m keeps both in frame for the first second or so; the numbers cover the rest. |
| - ±1° of gear play in every joint did not break either walk in simulation, even for the policies that never trained with it: level 5 held 6 of 7 standing takes with it, the same as its retrained twin. The extra 3,000 iterations with the play made no difference I can measure on walk-and-recover and only a small-sample one on the flat walk. What it does on a real servo is untested. |
| - A rough-terrain take proves nothing unless it is forced onto the rough part. Play builds a random terrain grid, every patch spawns the duck on a flat centre platform, and play shortens the push interval, so a default take is a push test on flat ground. Force the terrain row and type, switch pushes off and spawn the duck off the platform. |
| - The maker's fallen and height checks measure the trunk against the patch centre's height, which on pyramid stairs and slopes is the top platform. An upright duck a few steps down read as fallen and was reset 8 s later. Level 5c measures against the ground under the feet. |
| - The stock rough task never switches the terrain generator to curriculum mode, so its terrain levels were rows of random patches, not a difficulty ladder. Level 5c turns curriculum mode on; a sit-and-stand task cannot use the same ladder, because its level term promotes only an env that walked half a patch. |
| - Gear play changed nothing measurable on the get-up and sit-and-stand policies either: level 2b matched level 2 take for take, and level 3b's 4 of 5 against 3 of 5 is inside chance. |
| - A warm start reads low for the first few dozen iterations (every env begins a fresh episode at the resume), and level 5c logged ten iterations with huge negative mean rewards (lowest -1,402,121.9 at iteration 21,509) while the episode length stayed at about 660 steps. Not investigated yet; the charts clip these and say so, and the raw values are in the csv. |
| - The rough walking task resets a tipping duck at about 55 to 60° of tilt, before it lands, so a scorer that waits for a 60° tilt misses falls. Count the reset lines inside the window as falls, and check each one against the tilt (one reset at 3.7 m out with the duck level was the tile edge, not a fall). |
| - A rough-terrain retrain can buy nothing. Level 11d's mean reward never moved from its first 100 iterations, and its proof takes are worse than its seed's on rough ground, on flat ground and from the floor. The curve said so first; the takes confirmed it. |
| - Random spawns hide a regression or invent one. The first gear-play batch spawned each take in a random pose, and one policy drew back and belly spawns while the other drew standing ones; forcing the spawn (standing, face down, face up) and recording each policy in every pose made the comparison fair. |
| - Three to five takes per policy cannot separate a gain from noise. Level 1d's first set of five read as no gain (3 of 5 against 4 of 5 at normal grip); extending every condition to 13 to 15 takes showed the gain where the fork was aimed (slippery floor with shoves, 8 of 15 against 1 of 15) and still none at normal grip. Both sets are kept and pooled. |
| - A heavier trunk in training bought nothing: level 1b already walked with +60 g on its back in sim, so level 1c was no better in any condition. |
| - A take aimed at a floor feature needs a fixed walking command. Play samples random commands, and a level 1b control take on the step floor stayed within 0.73 m of the platform centre (edge at 0.75 m) until a reset, so it never met the step. Every 1e take walks straight ahead at 0.3 m/s. |
| - Keep a no-feature control on the same layout. Level 1e also held more takes than level 1b at 1 mm (10 of 15 against at most 7 of 15, two-sided p = 0.46), so some of its lead on the step may be general sturdiness under shoves rather than a step skill. Its training tracking error also rose (0.34 to 0.47 over the 100-iteration windows from the resume to the end). |
| - A generated floor turns on a termination a plane never fires. Level 1e's training log shows `out_of_terrain_bounds` at 0.01 per iteration over the last 100 iterations (`fell_over` 0.81), so the fork file's comment that a 20 s episode cannot reach the bounds is not exact; `nan_state` fired in one iteration of 3,000. |
| - A rival's low score on its own condition needs a re-run before it means anything. In 1h, walk-slip held 3 of 15 on the slippery-floor check against 8 of 15 in its own takes. Same checkpoint, same friction and shove; the batch differed in the fixed walking command and the generated floor. Re-running each change alone gave 7 of 15 and 6 of 15, so it was a low draw on the harder setup, not a regression, and the line reads as no loss rather than a win. |
| - Parallel take batches need their own temp folders. Two batches named their takes the same way and one deleted the other's video folder mid-take; the logs survived because they go to stdout, the clips did not. |
| - A verdict watcher can die without a trace. 1g and 1h finished training on a Sunday and sat four days with no proof takes, because the watcher that should have started them logged one line and vanished. |
| - Two forks that each won did not add up. Level 1i trained 1h's floor and 1g's hills together: it walked the hill as well as walk-slope (15 of 15 each) and stayed level with walk-mix on walk-mix's floor, but on the combined floor it held 5 of 15 against walk-mix's 3 of 15 (p = 0.34), and its reward was flat from the first 1,000 of its 3,000 iterations. The hill floor also costs about twice as much per iteration as the step floor (2.41 against 1.29 s/it). |
| - A proof batch cannot share the card with a job that books almost all of its memory. 1i's first batches started a minute after a bench job took 29 to 32 GB of the 32 GB card and most takes died on allocation errors; they were set aside and rerun unchanged on the idle card, and only the rerun is scored here. |
|
|
| ## Posts |
|
|
| Each level was posted the day it landed: [level 1](https://x.com/witcheer/status/2099877597622513821), [level 3](https://x.com/witcheer/status/2101249891791941866), [level 5](https://x.com/witcheer/status/2101760067401867627), [level 6](https://x.com/witcheer/status/2102076057411031472), [level 7](https://x.com/witcheer/status/2102652392453935248). |
|
|
| The real robot arrives in a few months. Until it does, every number in this repo is a simulation number. |
|
|
