Research map / Reinforcement learning / Visual and representation RL
Subequivariant Graph Reinforcement Learning in 3D Environments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 O_g(3)-subequivariant Set instead of SWAT | morphology-agnostic policy · 3D_CWHH++ cross-domain zero-shot, six held-out morphologies | return: 206.8±37.4 vs 17.9±13.7; 243.7±32.3 vs 114.9±40.3; 161.9±3.4 vs 152.0±6.8; 180.0±6.5 vs 156.6±1.7; 1078.1±722.8 vs 4.3±1 | Helped |
| Use Set instead of SWAT+HN | morphology-agnostic policy · 3D_CWHH++ cross-domain zero-shot, six held-out morphologies | return: 206.8±37.4 vs 26.3±72.4; 243.7±32.3 vs 156.8±11.1; 161.9±3.4 vs 130.2±2.1; 180.0±6.5 vs 152.9±36.8; 1078.1±722.8 vs 786. | Helped |
| Use equivariant Set rather than heading-normalized SWAT | policies tested with a 180° heading bias · 3d_humanoid_9_full and 3d_cheetah_14_full | return: 10106.4±2023.4 vs 92.3±17.8; 14957.9±758.0 vs -1.2±0.4 | Helped |
| Train on a fixed initial orientation and test at arbitrary orientations using Set rather than SWAT | 3d_cheetah_14_full · Fixed-orientation training, test at 180°, 1M training steps | return: 4756.3±103.4 vs -5.6±8.5 | Helped |
| Remove gravity from the geometric inputs and treat it as a scalar (ablation) | Set · 3d_humanoid_9_full | return | Hurt |
| Remove gravity from the geometric inputs and treat it as a scalar (ablation) | Set · 3d_cheetah_14_full | return | No clear effect |
| Remove the height feature (ablation) | Set · 3d_humanoid_9_full | return | Hurt |
| Remove the height feature (ablation) | Set · 3d_cheetah_14_full | return | No clear effect |
| Treat both gravity and target direction as scalars rather than directional vectors (ablation) | Set · 3d_humanoid_9_full and 3d_cheetah_14_full | return | Hurt |
| Treat all geometric vectors as scalar inputs, removing equivariant geometric processing (ablation) | Set · 3d_humanoid_9_full and 3d_cheetah_14_full | return | Hurt |
| Replace equivariant torque-vector projection with an invariant action mapping (ablation) | Set · 3d_humanoid_9_full and 3d_cheetah_14_full | return | Hurt |
Closest papers by meaning
- SiT: Symmetry-invariant Transformers for Generalisation in Reinforcement Learning (ICML 2024 · similarity 0.88)
- METRA: Scalable Unsupervised RL with Metric-Aware Abstraction (ICLR 2024 · similarity 0.87)
- Subequivariant Reinforcement Learning in 3D Multi-Entity Physical Environments (ICML 2024 · similarity 0.87)
- 3D Equivariant Visuomotor Policy Learning via Spherical Projection (NeurIPS 2025 · similarity 0.87)
- Geometry-aware RL for Manipulation of Varying Shapes and Deformable Objects (ICLR 2025 · similarity 0.87)
- Graph-Assisted Stitching for Offline Hierarchical Reinforcement Learning (ICML 2025 · similarity 0.86)
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.