HalfCheetah
State and action spaces
The default 17-value observation is qpos[1..] followed by all qvel; retaining root x produces 18 values. The default action is 6 controls in \([-1,1]\).
Dynamics, reward, and info
\[ v_x=\frac{x_{t+1}-x_t}{\Delta t},\qquad r=w_fv_x-w_u\lVert u\rVert_2^2. \]
Info contains x_position, x_velocity, reward_forward, and reward_ctrl. Reset adds uniform position noise and Gaussian velocity noise of configured scale.
Episode end
There is no state termination. The episode truncates at max_episode_steps.
Configuration
| Builder setting | Default | Meaning |
|---|---|---|
| XML | embedded model | Custom model source |
frame_skip | 5 | Positive number of MuJoCo steps per action |
| forward weight | 1 | Velocity reward coefficient |
| control weight | 0.1 | Squared-action penalty coefficient |
| reset noise | 0.1 | Non-negative initialization-noise scale |
| position exclusion | true | Omit root x from observation |
| horizon | 1000 | Truncation limit |
Reference: Gymnasium HalfCheetah.