legged_gym

Imported from yongkanghui75-a11y/LeggedGym-Ex (legged_gym/AGENTS.md). Install upstream with npx skills add yongkanghui75-a11y/LeggedGym-Ex --skill legged_gym. Copyright stays with the author.

LEGGED_GYM CORE FRAMEWORK

Core legged robot RL framework with multi-simulator support (Genesis, IsaacGym, IsaacLab).

STRUCTURE

KEY FILES

CONVENTIONS

Config Pattern: Nested classes with class env, class rewards, class commands, etc.

class MyRobotCfg(LeggedRobotCfg):
    class env:
        num_observations = 48
    class rewards:
        class scales:
            tracking_lin_vel = 1.0

Task Registration: Add to envs/__init__.py:

task_registry.register("robot_name", RobotClass, Cfg, CfgPPO)

Simulator Selection: export SIMULATOR=genesis|isaacgym|isaaclab

ANTI-PATTERNS

1. Observation Changes: Modifying obs_buf requires updating ALL _reward_* methods
2. IsaacGym Reset Bug: After reset(), call self.simulator.forward() before reading rigid body states
3. IsaacLab CPU Tensors: Domain randomization tensors must be on CPU (set_material_properties, set_masses, set_coms)
4. Genesis XML: Must provide XML file path when using Genesis simulator
5. Terrain Flags: Cannot use curriculum=True with selected=True simultaneously
6. IsaacLab Heightfield: Heightfield terrain not implemented for IsaacLabSimulator

PATTERNS

Type Aliases: ObsBuf = Tensor, Action = Tensor, Reward = Tensor in base classes

Config Validation: Extensive assertions in LeggedRobot.__init__() catch config errors early

Debug Flags: cfg.env.debug* enable visualization (height points, depth images, etc.)