关于Roboschool Half-Cheetah观测空间及状态定义的技术咨询

Hey Claire! Let's break down this confusion around the Roboschool Half-Cheetah's observation space and that comment in half_cheetah.xml—I've run into this exact head-scratcher before, so I know where you're coming from.

1. Why the Observation Space is 26D (Not 18D) #

You're probably missing a chunk of the observation that's easy to overlook. The full 26-dimensional space breaks down into three key parts:

• Root body state (4 dimensions): X-position, X-velocity, pitch angle (Z-axis rotation), and pitch angular velocity.
• Joint states (12 dimensions): Each of the 6 joints (bthigh, bshin, bfoot, fthigh, fshin, ffoot) contributes two values: their current angle and angular velocity.
• Limb linear velocities (10 dimensions): The remaining dimensions track X and Z axis linear velocities for each of the cheetah's non-root body segments. These are often overlooked when manually counting, which explains why you landed at 18 instead of 26.

2. The Persistently Zero Dimensions #

Those dimensions that stay 0 across all timesteps make total sense once you remember the Half-Cheetah is a planar robot—it only moves along the X-axis and pitches around the Z-axis. Any state tied to the Y-axis (perpendicular to the movement plane) like root Y-position, Y-velocity, or angular velocities around non-Z axes will never change from 0. These are included in the observation space for consistency with Mujoco's general robot model structure, even though they don't affect the robot's behavior.

3. What That XML Comment Means #

The comment in half_cheetah.xml:

Cheetah Model The state space is populated with joints in the order that they are defined in this file. The actuators also operate on joints.

This is a quick, practical reference for how the robot's state and controls map to the XML:

• Joint state order: The joint-related values in the observation array (angles and velocities) follow the exact order the joints are listed in the XML. If you check the <joint> tags, you'll see bthigh first, then bshin, bfoot, fthigh, fshin, ffoot—that's the same sequence in the observation.
• Actuator-joint mapping: The robot's actuators (the motors driving movement) are linked directly to these joints in the same order. So the first actuator controls bthigh, the second bshin, and so on.

内容的提问来源于stack exchange，提问作者Claire