Research map / Reinforcement learning / Visual and representation RL
Contrastive Learning as Goal-Conditioned Reinforcement Learning
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 contrastive RL with goal relabeling instead of prior offline baselines | Offline goal-conditioned RL · D4RL AntMaze large-play-v2 and large-diverse-v2 | normalized score: 41.6 {( ±6.0)} vs 39.6; 49.3 {( ±6.3)} vs 47.5 | Helped |
| Increase the number of critics from 2 to 5 (ablation) | Offline Contrastive RL + BC · D4RL AntMaze tasks | normalized score: 81.9 {( ±1.7)} -> 79.8 {( ±1.4)}; 75.4 {( ±3.5)} -> 77.6 {( ±2.8)}; 71.5 {( ±5.2)} -> 72.6 {( ±2.9)}; 72.5 {( ±2.8)} -> | Mixed |
| Add filtering of relabeled trajectories based on commanded-versus-reached-goal likelihood (ablation) | Contrastive RL (NCE) · Goal-conditioned RL tasks | success rate | Hurt |
| Sample future states rather than random states as actor-loss goals (ablation) | Contrastive RL (NCE) · Goal-conditioned RL tasks | success rate | Hurt |
| Use a mixture of future goals and random goals for the actor loss (ablation) | Contrastive RL (NCE) · Goal-conditioned RL tasks | success rate | Mixed |
| Replace the NCE objective with the CPC objective (ablation) | Contrastive RL · Goal-conditioned tasks | success rate | Mixed |
| Replace NCE with C-learning (ablation) | Contrastive RL · Goal-conditioned tasks | success rate | Mixed |
| Combine the NCE objective with C-learning (ablation) | Contrastive RL · Goal-conditioned tasks | success rate | Helped |
| Add image data augmentation and auxiliary representation losses to TD3+HER (ablation) | TD3+HER · fetch reach and sawyer push image-based tasks | success rate | Helped |
| Use Contrastive RL (NCE) instead of TD3+HER with representation-learning additions | Image-based goal-conditioned RL · Image-based tasks | success rate | Helped |
| Normalize the final representation, add a final activation, or use a learnable temperature (ablation) | Contrastive RL (NCE) · Goal-conditioned tasks | success rate | Hurt |
| Use contrastive RL instead of a forward model as goal dimension increases | sawyer push · Goal dimensions 4 through 7 | Euclidean distance to goal | Helped |
Closest papers by meaning
- Offline Goal-conditioned Reinforcement Learning with Quasimetric Representations (NeurIPS 2025 · similarity 0.90)
- Learning Goal-Conditioned Representations for Language Reward Models (NeurIPS 2024 · similarity 0.89)
- Representation-Based Robustness in Goal-Conditioned Reinforcement Learning (AAAI 2024 · similarity 0.89)
- ContraDiff: Planning Towards High Return States via Contrastive Learning (ICLR 2025 · similarity 0.87)
- Stabilizing Contrastive RL: Techniques for Robotic Goal Reaching from Offline Data (ICLR 2024 · similarity 0.87)
- Towards Principled Representation Learning from Videos for Reinforcement Learning (ICLR 2024 · 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.