Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Goal-Conditioned Agents that Learn Everything All at Once
Michael Matthews, Matthew Thomas Jackson, Michael Beukman, Thomas Foster, Alistair Letcher, Scott Fujimoto, Cédric Colas, Jakob Nicolaus Foerster
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 |
|---|---|---|---|
| learn all goals jointly with LEO output heads instead of naive all-goals relabelling | PQN · CraftaxGC (512 goals) | throughput: 34% slowdown vs single-goal; naive relabelling 264x slower than LEO | Improved |
| add HER relabelling | PQN · Craftax-Classic | mean success rate | Improved |
| add HER relabelling | PQN · Craftax (512 goals) | mean success rate | Improved |
| replace PQN with LEO all-goals updates | PQN · Craftax (512 goals) | mean success rate | Improved |
| replace PQN with LEO all-goals updates | PQN · Craftax-Classic | mean success rate | Mixed |
| use LEO alone as actor on easy goals | LEO · Craftax selected easy goals | success rate | Hurt |
| add a UVFA network alongside LEO (Dual LEO PQN) | LEO · Craftax (512 goals) | mean success rate | Improved |
| distill LEO into PPO actor-critic (Dual LEO PPO) | PPO · Craftax (512 goals) | mean success rate | Improved |
| add LEO teacher to PPO | PPO · Craftax-Classic | mean success rate | Improved |
| add goal curriculum of previously seen goals (ablation) | LEO and Dual LEO · Craftax | mean success rate | Hurt |
| use SAC+LEO instead of SAC+HER/CRL | SAC · JaxGCRL ant maze U Maze | success rate | Improved |
| use SAC+LEO instead of SAC+HER/CRL | SAC · JaxGCRL ant maze (larger maze) | success rate | No clear effect |
Closest papers by meaning
- Accelerating Goal-Conditioned Reinforcement Learning Algorithms and Research (ICLR 2025 · similarity 0.87)
- Discrete Compositional Representations as an Abstraction for Goal Conditioned Reinforcement Learning (NeurIPS 2022 · similarity 0.87)
- Visual Reinforcement Learning with Imagined Goals (NeurIPS 2018 · similarity 0.86)
- Flattening Hierarchies with Policy Bootstrapping (NeurIPS 2025 · similarity 0.86)
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.85)
- Improving the Continuity of Goal-Achievement Ability via Policy Self-Regularization for Goal-Conditioned Reinforcement Learning (ICML 2025 · similarity 0.85)
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.
