Principle:Haosulab ManiSkill Action Space Normalization
| Knowledge Sources | |
|---|---|
| Domains | Robotics, Reinforcement_Learning, Control_Theory |
| Last Updated | 2026-02-15 08:00 GMT |
Overview
Controllers optionally normalize actions from a bounded [-1, 1] range to the actual joint limits or configured bounds, ensuring that learning algorithms receive well-conditioned action spaces regardless of the underlying physical units.
Description
The Action Space Normalization principle addresses the mismatch between what reinforcement learning and imitation learning algorithms expect (bounded, zero-centered action spaces) and what physics engines require (actions in physical units like radians, radians/second, or meters). Each controller in ManiSkill can be configured with normalize_action=True, which maps the policy's [-1, 1] output to the controller's configured bounds via linear interpolation, and clips the result to valid ranges.
This normalization is critical because different joints have vastly different physical ranges (e.g., a revolute joint may span [-2.8, 2.8] radians while a gripper joint spans [0, 0.04] meters). Without normalization, the policy must learn these heterogeneous scales implicitly, which slows training. With normalization, every action dimension has the same semantic scale.
Usage
This principle applies whenever:
- RL or IL policies output actions that need to be mapped to physical joint ranges.
- A robot has joints with heterogeneous ranges that would create poorly conditioned optimization landscapes if exposed directly.
- Action clipping is needed to prevent physically impossible joint targets.
Theoretical Basis
Linear Mapping: Given action a ∈ [-1, 1] and bounds [low, high], the physical action is: physical = low + (a + 1) / 2 * (high - low).
Clipping: After denormalization, actions are clipped to [low, high] to prevent out-of-range drive targets.
Per-Controller Configuration: Each controller independently decides whether to normalize. Position controllers typically normalize to joint limits; velocity controllers normalize to velocity bounds; EE pose controllers normalize to workspace bounds.
Related Pages
- Implementation:Haosulab_ManiSkill_BaseController -- Base controller implementing normalize_action flag and action space construction.