Legged Locomotion RL — Sim-to-Real

I have generated the reward structure and the corresponding environment observation space for the quadruped trot task. The rewards use exponential kernels for tracking and squared penalties for hardware longevity.

Quadruped Trot Reward Configuration

# Environment Observation Baseline
# Total Dimension: 48 (Proprioceptive)
obs_spec = {
    "base_ang_vel": 3,        # Angular velocity in base frame
    "projected_gravity": 3,   # Gravity vector in base frame
    "commands": 3,            # [vx, vy, yaw_rate]
    "joint_pos": 12,          # Relative to default pose
    "joint_vel": 12,          # Current joint velocities
    "last_actions": 12        # Previous step actions for smoothness
}

# Domain Randomization Ranges
randomization = {
    "friction": [0.5, 1.25],
    "added_mass": [-1.0, 1.0], # kg
    "pd_gains": [0.85, 1.15],  # Scaling factor
    "latency": [0.0, 0.02]     # up to 20ms
}

Next steps

1. Integrate these kernels into your LeggedRobotCfg reward class.
2. Run python scripts/train.py and monitor the episode_sum_lin_vel in Tensorboard.
3. Validate the export with deploy_config_check.py before moving to the harness.