witcheer/microduck-skill-tree

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 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 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.

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 @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 @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; no rough-ground take recorded yet	microduck-walk-rough
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 @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
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
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 @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 @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 @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
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 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). 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 and 5c 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-*). Verdicts are read from these numbers, not from the video.
• patches/*.py (level 5c only): the local changes to the maker's code the level trained with.
• 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, 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 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/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.
• 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.

Posts

Each level was posted the day it landed: level 1, level 3, level 5, level 6, level 7.

The real robot arrives in a few months. Until it does, every number in this repo is a simulation number.