Research map / Reinforcement learning / Visual and representation RL
Simple Emergent Action Representations from Multi-Task Policy Training
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 |
|---|---|---|---|
| multi-task policy with shared sensory and task embeddings (EAR-SAC) vs one-hot embedding policy (OHE-SAC) (ablation) | SAC policy with latent task embeddings · HalfCheetah-Vel, Ant-Dir, Hopper-Vel, Walker-Vel | final return | Helped |
| add unit-sphere normalization of LTEs and LSEs (ablation) | EAR-SAC · Ant-Dir interpolation | mid-task and 1:1 interpolation success rate: 64->70 of 72 (mid-task); 59->66 of 72 (1:1) | Helped |
| add random noise injection to LTEs during training (ablation) | EAR-SAC · Ant-Dir training | tasks with trajectory failures: 16 of 72 failures -> 0 of 72 | Helped |
| add random noise injection to LTEs during training (ablation) | EAR-SAC · Ant-Dir training curves | training performance | No clear effect |
| EAR-SAC latent action representation vs MHMT-SAC and meta RL baselines | SAC with LTE interface · Ant-Dir final return | return (last epoch): 1216.7 vs 1020.0 (MHMT-SAC) | Helped |
| EAR-SAC vs MHMT-SAC on HalfCheetah-Vel final return | SAC with LTE interface · HalfCheetah-Vel | return (last epoch): -136.9 vs -145.2 (MHMT-SAC) | Mixed |
| EAR-SAC vs MHMT-SAC on Hopper-Vel and Walker-Vel final return | SAC with LTE interface · Hopper-Vel, Walker-Vel | return (last epoch): 173.5 vs 172.9 (Hopper); 172.7 vs 160.6 (Walker) | No clear effect |
| EAR-SAC interpolation of task embeddings for intra-action interpolation | EAR-SAC · HalfCheetah-Vel, Ant-Dir, Hopper-Vel, Walker-Vel | mid-task interpolation success: 26/27, 70/72, 27/27, 27/27 | Helped |
Closest papers by meaning
- Learning Action Representations for Reinforcement Learning (ICML 2019 · similarity 0.90)
- The Natural Language of Actions (ICML 2019 · similarity 0.89)
- Act to See, See to Act: Diffusion-Driven Perception-Action Interplay for Adaptive Policies (NeurIPS 2025 · similarity 0.88)
- Representation Learning Enables Scalable Multitask Deep Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
- Universal Humanoid Motion Representations for Physics-Based Control (ICLR 2024 · similarity 0.86)
- Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (NeurIPS 2024 · 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.