Training MicroDuck

IsaacContrib-Velocity-Flat-MicroDuck and IsaacContrib-Velocity-Rough-MicroDuck train MicroDuck to walk with Newton MJWarp and BAM servos. Add presets=backlash to select the robot with a passive play hinge in series with each servo.

Walking recipe

Both tasks use the AGILE-inspired combined recipe by default. It tracks horizontal root-link velocity in the heading frame with a tolerance of 0.15 m/s and world-frame yaw rate with a tolerance of 0.35 rad/s, each with reward weight 4.0. The nominal joint-posture reward applies only to standing commands, and the action-rate penalty increases from -0.1 to -1.0 between training iterations 3,000 and 7,000. The actor and critic have 61 and 76 inputs, respectively, and use normalized MLPs with hidden layers [512, 256, 128] and ELU activation.

For example, train on rough terrain with backlash:

uv run isaaclab train --task IsaacContrib-Velocity-Rough-MicroDuck \
    --rl_library rsl_rl presets=backlash

Use IsaacContrib-Velocity-Flat-MicroDuck for flat ground, or omit backlash for the regular robot.

Rough terrain advances after an episode survives and tracks its commands, with at least 1 m of commanded and actual walking and 5 s of translation commands. It accumulates path length, so turns and changes of direction retain progress. Mostly stationary episodes hold their difficulty; failures and poor tracking lower it.

Distributed training

With 4,096 environments per GPU, MicroDuck preserves the single-GPU sample budget, rounded up to full rollouts, by automatically scaling with global WORLD_SIZE (default: 1). This also handles nodes with unequal worker counts.

uv run isaaclab train_multigpu --num_gpus 8 \
    --task IsaacContrib-Velocity-Rough-MicroDuck --rl_library rsl_rl \
    --num_envs 4096 --seed 42 presets=backlash

For N ranks, training uses ceil(50000 / N) iterations and a ceil(250 / N) checkpoint interval. All six fixed-step curriculum boundaries are divided by N, rounding up. Explicit iteration, checkpoint and curriculum overrides take precedence. To override N, supply both env.ngpus (GPUs per node) and env.nnodes; their product affects recipe scaling only. Both default to None; incomplete pairs raise an error.

Select actual workers with --num_gpus (per node) and --nnodes; see Train across multiple nodes. Single-GPU defaults, the 24-step rollout and PPO settings are unchanged. Terrain progression remains episode-based. Larger optimizer batches can change convergence despite the matching sample budget.

Previous walking recipe

To retain the previous recipe in a custom task config, use mdp.track_linear_velocity with weight 2.0 and std=math.sqrt(0.1), and mdp.track_angular_velocity with weight 2.0 and std=math.sqrt(0.5). Set rewards.pose.params["standing_only"] = False and restore the action-rate weight stages to (500, -0.2), (750, -0.4), (1000, -0.6), (1250, -0.8), and (1500, -1.0).