Cheetah

A planar bipedal cheetah-style runner. The agent torques 6 actuators across the body to maintain forward locomotion across flat ground.

CheetahRun

Property	Value
Canonical ID	mjx/cheetah_run-v0
Action space	Box(-1.0, 1.0, (6,), float32)
Observation space	Box(-inf, inf, (17,), float32)
Episode length	1000
Config	{"ctrl_dt": 0.01, "sim_dt": 0.01, "naconmax": 100_000, "njmax": 100}

Description

The half-cheetah body runs rightward across flat ground. Six actuators across the torso and legs drive the gait. The body is planar, so the cheetah can't tip sideways, but it can collapse forward or backward — which is the main failure mode for naive policies that throw all their weight into raw acceleration.

Rewards

Uses a dense reward with a tolerance indicator over forward speed:

speed = torso_subtreelinvel[0]  # forward velocity
reward = tolerance(
    speed,
    bounds=(RUN_SPEED, inf),
    margin=RUN_SPEED,
    value_at_margin=0,
    sigmoid="linear",
)

tolerance is DM Control's smooth indicator. Here it's used with a linear sigmoid, so the reward returns:

• 1.0 once speed >= RUN_SPEED.
• A linear ramp from 0.0 (at speed 0) to 1.0 (at RUN_SPEED).
• 0.0 if speed dips negative (running backwards).

Starting state

obs = [-0.1     0.0349 -0.0361  0.0071 -0.0729  0.0074 -0.0493 -0.0379
       -0.0113 -0.0086  0.0134 -0.0104  0.0007 -0.0228 -0.0005 -0.0484 -0.0615]

(joint positions followed by joint velocities — body at rest with a small randomisation.)

Termination

Episode ends when step >= max_steps (default 1000). No early termination on falling.

Usage

import envrax
env = envrax.make("mjx/cheetah_run-v0")

Reference

Upstream: mujoco_playground/_src/dm_control_suite/cheetah.py.