Research map / Reinforcement learning / Visual and representation RL
A Primer on SO(3) Action Representations in Deep Reinforcement Learning
Martin Schuck, Sherif Samy, Angela P. Schoellig
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use delta tangent-space actions in the local frame instead of global matrix actions | PPO/SAC/TD3 · Idealized SO(3) rotation environment | episode return: SAC dense -4.7 -> -2.9; SAC sparse -29.4 -> -7.9 | Improved |
| scale tangent increments to the permissible rotation range (ablation) | delta tangent actions · Idealized SO(3) rotation environment (dense) | episode return: about -1.5 difference without scaling | Improved |
| add unit-rotation centering to the policy mean for delta quaternion/matrix actions (ablation) | PPO · Idealized SO(3) rotation environment | learning curve | Improved |
| add unit-rotation centering to the policy mean for delta quaternion/matrix actions (ablation) | SAC and TD3 · Idealized SO(3) rotation environment | learning curve | Mixed |
| project sampled actions onto the manifold before computing probabilities (ablation) | PPO with quaternion/matrix actions · Idealized SO(3) rotation environment (dense) | learning curve | Hurt |
| project sampled actions onto the manifold (ablation) | SAC with quaternion/matrix actions · Idealized SO(3) rotation environment (dense) | learning curve | No clear effect |
| use quaternion actions (double cover) instead of rotation matrices | global actions · Idealized SO(3) rotation environment | episode return: PPO -5.4 -> -11.5 | Hurt |
| increase entropy coefficient (ablation) | PPO/SAC with tangent, quaternion, matrix actions · Idealized SO(3) rotation environment | final policy | No clear effect |
| use SAC with entropy maximization and off-manifold actions under sparse rewards | global matrix actions · Idealized SO(3) rotation environment (sparse) | episode return: TD3 -6.4 vs SAC -29.4 | Hurt |
| use delta tangent actions on drone control with PPO | PPO · Drone trajectory tracking and racing | reward | Improved |
| use delta tangent actions instead of other representations | TD3+HER · PickAndPlaceOrient | success rate: 54.1% -> 69.8% | Improved |
| use delta tangent actions instead of quaternions | SAC · RoboSuite (9 tasks) | reward | No clear effect |
Closest papers by meaning
- Learning Action Representations for Reinforcement Learning (ICML 2019 · similarity 0.88)
- Simple Emergent Action Representations from Multi-Task Policy Training (ICLR 2023 · similarity 0.88)
- For SALE: State-Action Representation Learning for Deep Reinforcement Learning (NeurIPS 2023 · similarity 0.88)
- Geometry of Neural Reinforcement Learning in Continuous State and Action Spaces (ICLR 2025 · similarity 0.87)
- Beyond Distributions: Geometric Action Control for Continuous Reinforcement Learning (ICLR 2026 · similarity 0.87)
- Demystifying Action Space Design for Robotic Manipulation Policies (ICML 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
