microduck-running
A robustified forward-running policy for the stock MicroDuck on flat ground. The default artifact is iteration 12,195, continued from the 1.687 m/s speed frontier with progressively stronger velocity pushes, centre-of-mass variation, and initial tilt.
Status: simulation-only hardware candidate; not validated on hardware. It retains 1.651 m/s mean body-forward speed in the nominal simulation battery and 1.612 m/s under the combined backlash-plus-disturbance battery. Heading and lateral drift remain substantial.
Normal simulation rollout β iteration 12,195
Run-into-mat edit β legacy iteration 8,749
Open the full silent run-into-mat MP4.
The mat video uses the preserved iteration-8,749 controller, not the new root policy. Its obstacle, collision augmentation, camera treatment, post-impact motor cutoff, and sparks are scripted demonstration elements. The policy has no obstacle perception and did not learn to react to a mat. Both videos are silent.
Versions
| Role | Iteration | Files | Note |
|---|---|---|---|
| Default robust candidate | 12,195 | policy.onnx, checkpoint.pt |
Recommended starting point for further simulation and cautious hardware research |
| Speed-focused parent | 11,748 | lineage/iteration-11748/ |
1.687 m/s frontier used to start robustification |
| Legacy mat-video policy | 8,749 | legacy/iteration-8749/ |
Exact controller used in the run-into-mat edit |
Command
The 13-D command block is [twist(3), head_pose(4), body_pose(6)].
| Slot | Meaning |
|---|---|
twist[0] |
forward-velocity command; trained and evaluated through 2.20 m/s |
twist[1] |
lateral command; use 0 (only Β±0.02 m/s was seen in training) |
twist[2] |
yaw-rate command; use 0 (only Β±0.05 rad/s was seen in training) |
head_pose[0:4] |
unused; zeros |
body_pose[0:6] |
unused; zeros |
An exact-zero standing bucket made up 3% of training environments, so
twist = [0, 0, 0] is the intended idle command. The speed objective rewards
forward progress and can exceed the command; twist[0] is not a hard speed
limiter.
Try it in simulation
The matching source is
Vottivott/microduck-playground@828d950:
git clone https://github.com/Vottivott/microduck-playground.git
cd microduck-playground
git checkout 828d950134e29a8d04cbb51720a22c8729047fb7
uv sync
hf download HannesVonEssen/microduck-running policy.onnx config.json --local-dir policies/running
uv run scripts/infer_policy.py \
--walking policies/running/policy.onnx \
--new-cmd-obs \
--lin-vel-x 2.2
Continue training
checkpoint.pt is the complete iteration-12,195 PPO state: actor, critic,
optimizer, observation normalizers, and curriculum counter.
mkdir -p logs/rsl_rl/running/release-12195
cp checkpoint.pt logs/rsl_rl/running/release-12195/model_12195.pt
MICRODUCK_RUNNING_TARGET_MAX_SPEED=2.2 \
MICRODUCK_RUNNING_SPEED_CAP=2.4 \
MICRODUCK_RUNNING_HIGH_SPEED_STAGE_INTERVAL=750 \
MICRODUCK_RUNNING_ACTION_RATE_WEIGHT=-0.10 \
MICRODUCK_RUNNING_FORWARD_PROGRESS_WEIGHT=5.0 \
MICRODUCK_RUNNING_ROBUST_PUSH_MPS=0.10 \
MICRODUCK_RUNNING_ROBUST_TRUNK_COM_M=0.008 \
MICRODUCK_RUNNING_ROBUST_HEAD_COM_M=0.006 \
MICRODUCK_RUNNING_ROBUST_INITIAL_TILT_DEG=2.0 \
uv run train Mjlab-Running-Flat-MicroDuck \
--agent.resume True --agent.load-run release-12195 \
--agent.load-checkpoint model_12195.pt --agent.max-iterations 100
PyTorch checkpoints use pickle internally; load them only from a repository and revision you trust.
Contract
- input:
obs, float32[1, 61] - output:
actions, float32[1, 14] - control rate: 50 Hz
- action scale: 1.0, joint-position targets around MicroDuck HOME
- observation normalizer: baked into
policy.onnx - output clipping: none; training also used unclipped actor outputs
- entry pose: standing
- robot: MicroDuck hardware revision 1, 14 Dynamixel XL330 servos
- hardware modifications: none
Checkpoint-to-ONNX parity over zero and deterministic random observations had
a maximum absolute action error of 7.63e-6. See
eval/checkpoint-onnx-parity.json.
Simulation evaluation
All batteries used a 2.20 m/s command, 1-second warm-up, and 10-second measurement horizon.
| Evaluation | Environments | Body-forward speed | Survival |
|---|---|---|---|
| Ordinary model, nominal | 512 | 1.651 m/s | 99.22% |
| Ordinary model, push/CoM/tilt stress | 512 | 1.635 m/s | 98.83% |
| Backlash model, nominal | 256 | 1.636 m/s | 98.83% |
| Backlash model, push/CoM/tilt stress | 256 | 1.612 m/s | 98.44% |
| High grip plus full stress | 512 | 1.642 m/s | 96.68% |
Stress means Β±0.10 m/s planar pushes every 3β6 seconds, Β±10 mm trunk CoM,
Β±6 mm head CoM, and Β±2Β° initial pitch/roll. The high-grip case widens foot
friction from 0.7β1.3 to 0.7β1.8. Raw outputs are under eval/.
The candidate was trained with the ordinary robot model, then evaluated unchanged with the backlash model. A separately backlash-trained continuation performed worse and was rejected. Robustification costs about 2.1% nominal speed relative to iteration 11,748 but gives the stronger overall release.
Sim-to-real boundary
The policy retains substantial deployment-oriented modeling:
- BAM XL330 M6 voltage control, back-EMF, current limiting, 6.5β8.2 V battery variation, and load-dependent sag;
- 3β6 physics-step actuator delay, 0β1-step IMU delay, and one-control-step joint-velocity lag;
- foot friction randomized from 0.7β1.3;
- trunk mass/inertia Β±5%, armature Β±10%, BAM joint friction Β±10%, encoder bias Β±0.015 rad, IMU misalignment up to 6Β°, and actor-observation noise;
- in the final stage: Β±0.10 m/s pushes, Β±8 mm trunk CoM, Β±6 mm head CoM, and Β±2Β° initial pitch/roll.
Important remaining relaxations are flat terrain, ordinary non-backlash training, disabled head/body pose tracking, weaker posture and angular-motion regularization, and no motor-gain randomization. Backlash was tested, not part of the selected policy's training distribution. Transfer is plausible, not proven; heading drift, impact loading, real contact, servo thermal limits, and unmodeled compliance remain concerns.
Hardware safety
For a first robot test, use a support/catch rig, keep a remote torque-off path, start with low commands and ramp conservatively, and log servo current, temperature, voltage, fall state, and commanded/applied twist. Do not begin at 2.20 m/s. Stop on repeated saturation, overheating, unexpected oscillation, or systematic forward pitching.
Provenance and files
Training task Mjlab-Running-Flat-MicroDuck; PPO with rsl_rl; checkpoint
iteration 12,195. Source:
Vottivott/microduck-playground@828d950,
based on
pollen-robotics/microduck_rl@d424a0c.
Exported with scripts/export.py, which bakes the observation normalizer into
the ONNX graph.
policy.onnxβ default robust deployment policycheckpoint.ptβ default resumable PPO checkpointmanifest.jsonβ machine-readable contract, evaluation, and provenanceeval/β raw evaluation records and checkpoint/ONNX parity recordmedia/preview.mp4β robust policy's 10-second close rolloutmedia/run-into-mat.mp4β scripted legacy iteration-8,749 demonstrationlineage/iteration-11748/β speed-parent model, checkpoint, video, and evalslegacy/iteration-8749/β exact mat-video policy, checkpoint, and clean videoSHA256SUMSβ artifact integrity hashes
Format inspired by
RemiFabre/microduck-flamingo-cycle.
- Downloads last month
- 73