robots.md

Robots

The benchmark's four MuJoCo Playground quadrupeds (open serial chains) next to Wojtek (closed-chain legs, environment.md). Select with robot=<name> in every entry point with a robot group, ROBOT=<name> in slurm/; sizes and the robot list: overview.md.

robot=	class	model	SCALE	standing trunk (m)	DR
go1	Go1Terrain	Playground go1_mjx_feetonly.xml	2.5	0.258	Playground go1/randomize.py
barkour	BarkourTerrain	menagerie google_barkour_vb/barkour_vb_mjx.xml	2.5	0.253	none (Playground has none)
spot	SpotTerrain	Playground spot_mjx_feetonly.xml	4.3	0.433	none (Playground has none)
go2	Go2Terrain	menagerie unitree_go2/go2_mjx.xml, rewritten feet-only	2.7	0.270	Go1's randomize.py

Shared (robots/terrain.py TerrainRobot, a mixin over the Playground env; a robot file keeps its constants, scene and what differs):

• Playground's own reward and obs, each robot's own PD gains (table below), feet-only collisions, blind actor.

• jax backend only (playground_backend): backend=auto resolves to jax, backend=warp is refused because foot contacts with the tiled arena's hfield diverge on warp far from the origin.

• The tiled ACCEL arena written by arena_scene (no asset files): one hfield geom floor (geom 0, so a friction DR keyed on geom 0 hits it), sized by HFieldConfig; reset spawns on spawn_xy, the tile centre FunctionalEnv sets.

• Commands: the forward slice of each robot's own Playground command box (native_command_range): vx uniform from 0 to the robot's max, vy and wz uniform over their native ranges, no standing commands; sample_command redraws from the same box on the env's own schedule.

  robot	vx (m/s)	vy (m/s)	wz (rad/s)	source
  Go1, Go2	0-1.5	+-0.8	+-1.2	go1/joystick.py command_config.a
  Barkour	0-1.5	+-0.8	+-0.7	barkour/joystick.py lin_vel_x/y, ang_vel_yaw
  Spot	0-1.0	+-0.8	+-1.0	spot/joystick.py command_config

• Eval commands (make_env(..., eval_travel=), used by brax_ued's eval, eval_brax and eval_sim2sim): vx only from EVAL_SPEED_MARGIN (1.2) x threshold / episode time up to the max, so every eval command can reach the success threshold (below) in the robot's level: Go1, Barkour 0.5-1.5, Go2 0.54-1.5, Spot 0.33-1.0 m/s at the benchmark's min_travel=1.0, level_steps=300. Training draws the full forward slice, so train success tops out below 1.

• Success threshold: robots.scale.min_travel(cfg.robot, cfg.min_travel) = min_travel x the robot's size, the one source for training success, the ACCEL / SFL / PLR scores, eval, sim2sim, domination and paper (both rebuild success from raw travel) and the export README.

• Level length: robots.scale.level_steps(cfg.robot, cfg.level_steps) = level_steps x the robot's EPISODE_SCALE control steps, the one source for training (unrolls per iteration, env-step count), eval, sim2sim and the export README; level_steps is in Wojtek's steps. EPISODE_SCALE is 1 except Spot's 2.6 (780 steps, 15.6 s): Wojtek's top vx per unit size (0.6 m/s / 0.1 m) over Spot's (1.0 m/s / 0.43 m), rounded to 0.1, so the 4.3 m threshold is the same share of the distance at top speed as for Wojtek, Go1 and Barkour. robot.scale does not change it.

  robot	threshold (m)	steps	/ distance at eval vx floor	/ distance at vx max
  Wojtek	1.0	300	0.83	0.28
  Go1	2.5	300	0.83	0.28
  Go2	2.7	300	0.83	0.30
  Barkour	2.5	300	0.83	0.28
  Spot	4.3	780	0.83	0.28

  Spot redraws its command every 200 steps (Barkour 500, Go1 / Go2 every 5 s on average); with turning, perfect tracking of eval commands reaches the threshold in 62 % of Spot's levels, 70 % of Go1's and Go2's, 95 % of Barkour's.

• forward_vel (train and eval logs) is measured along the commands' x axis: FunctionalEnv.forward_velocity, for these robots the local_linvel velocimeter the reward tracks (command_frame_velocity). Barkour's velocimeter site imu_frame faces the torso's -x, so the root body frame would log its forward walking as negative.

• Wojtek's fall rule at the robot's size: trunk below 0.6 x STANDING_HEIGHT above the terrain under it (terrain_height, read off hfield_data), or gravity z > -0.4 in the trunk frame.

• The critic gets a clean 5x5 height scan (Wojtek's grid x SCALE) appended to privileged_state.

• SCALE = standing trunk height / Wojtek's 0.10 m, rounded to 0.1; robot.scale overrides it (below). The trunk height is measured: the settled qpos[2] over flat ground under zero action (the PD pose), which sags below the kinematic height. tests/robots/test_terrain_robots.py re-measures it per robot. SCALE scales the arena (hfield_config, reach included), level lengths and min_travel; slopes do not scale.

• Needs mujoco_menagerie under Playground's external_deps (meshes); ensure_menagerie_exists() clones it at the pinned commit (setup.md); on a cluster whose login node cannot import mujoco, clone it on a compute node or by hand.

• eval_sim2sim's full_collision variant exists for Go1 and Spot only (FULL_COLLISION); Barkour has no full-collision model and Go2's is not wired, so those two refuse it with an error.

Overrides: gains and scale

RobotConfig (config/schema.py) has four fields that default to null (= the robot's own value), so conf/robot/<name>.yaml sets only name; a number replaces the robot's own value (robot.kp=50 robot.scale=2.0). kp must be > 0, kd and damping >= 0:

field	sets	Go1	Go2	Barkour	Spot
kp	actuator position gain (gainprm[0], biasprm[1] = -kp)	35	100	50	300
kd	actuator velocity gain kv (biasprm[2] = -kd)	0	0.707	0.5	20
damping	passive damping of the 12 leg joints (dof_damping[6:]; Playground's Kd for Go1 / Spot)	0.5	2.828	0.024	1
scale	size relative to Wojtek (SCALE)	2.5	2.7	2.5	4.3

• Gains: PDGains (robots/terrain.py) edits the compiled MjModel after the robot's own setup and rebuilds the MJX model; all-null leaves the model untouched. Exported kd (export/contract.py) = kd + damping.
• Scale: robots.scale.robot_scale(cfg.robot) is the one value read by hfield_config (arena lengths, reach), min_travel, the param level space (robot_scale), the eval suite, the critic's height scan and the travel bounds; make_env passes it to the robot. Pass the same overrides to train, eval, sim2sim, render and export of a run (ROBOT_ARGS in slurm/). Fall rule, spawn height and the level length (EPISODE_SCALE) do not scale (they are the robot's own).
• Wojtek refuses every non-null field.
• Defaults: Go1 and Spot keep Playground's gains, Barkour its menagerie model's (Playground's env would set kp 35, damping 0.5239), Go2 its own GAINS (robots/go2.py, below).
• Record: training writes <name>.robot.yaml beside the checkpoint (name, the compiled kp / kd / damping, scale, min_travel_m, level_steps (resolved); robots/record.py). eval and sim2sim rebuild it from their own config and refuse a mismatch, naming the fields; without the file they warn and run.

Per robot

• Go1: Playground Go1JoystickRoughTerrain, spawning 0.35 m above the tile.
• Barkour vB: Playground barkour.Joystick (18 dof, 12 joints clipped to its limits, perturbation kicks always on). Its env builds from a fixed menagerie scene and returns one obs vector (a 15-step history, 465 values), so BarkourTerrain builds the arena scene itself, adds the foot contact sensors in XML (Playground adds them with MjSpec), uses the vector as the actor obs and [obs, height scan] as the critic obs, and carries a copy of step whose termination is the shared fall rule plus Playground's joint-limit check: its own torso z < 0.18 is a world height, wrong on a pit or raised tile. Commands are redrawn every 500 steps.
• Spot: Playground spot.Joystick (Kp 300, Kd 1, 3-step joint-error history, feet position obs). get_gravity there is the up vector, so the fall rule negates it. Playground's early_termination flag and gravity z < 0.85 termination are replaced by the shared rule. Spot's standing trunk is 0.433 m, scale 4.3, reach 43 m. Commands are redrawn every 200 steps; levels are 780 steps (EPISODE_SCALE 2.6, above).
• Go2: Playground has no Go2 env: Go2Terrain is Go1Terrain (Go1's joystick reward and obs) on the menagerie model with GAINS kp 100, kv 0.707, joint damping 2.828: menagerie's kp 50 doubled, kv and damping x sqrt(2), so the damping ratio stays menagerie's. The trunk stands at 0.270 m (Go1's Kp 35 sags it to 0.219 m). go2_mjx.xml collides every link and lacks the names Go1's env reads, so feet_only_model rewrites it (each substitution must match exactly once): trunk base -> trunk, foot sites FL_foot -> FL, only the four foot spheres collide (no hfield-sphere margin in MJX), the upvector, local_linvel, per-foot *_global_linvel and *_floor_found sensors are added, the model's own keyframe is dropped. Go1's randomization fits (hfield = geom 0, trunk = body 1, 12 joints = dofs 6:).