Research map / Reinforcement learning / Visual and representation RL
Curious Representation Learning for Embodied Intelligence
Yilun Du, Chuang Gan, Phillip Isola
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 |
|---|---|---|---|
| freeze pretrained visual weights during downstream RL (ablation) | CRL representation · ObjectNav (Matterport3D) | Soft SPL | Improved |
| freeze convolutional weights of the from-scratch RL policy (ablation) | From Scratch policy · ImageNav/ObjectNav | Soft SPL: a fall of over 0.07 SoftSPL | Hurt |
| replace random exploration with CRL curiosity policy for gathering contrastive data (ablation) | contrastive learning · ImageNav (Gibson) | SPL: 0.0285 -> 0.0324 | Improved |
| replace random exploration with CRL curiosity policy for gathering contrastive data (ablation) | contrastive learning · ObjectNav (Matterport3D) | SPL: 0.0042 -> 0.0144 | Improved |
| replace learned-counts exploration with CRL (ablation) | contrastive learning · ObjectNav (Matterport3D) | Soft SPL: 0.110 -> 0.119 | Improved |
| replace RND representation with contrastive CRL | representation learning · ImageNav (Gibson) | Soft SPL: 0.124 -> 0.219 | Improved |
| use CRL representation instead of ImageNet initialization | frozen encoder · ImageNav (Gibson) | SPL: 0.0193 -> 0.0324 | Improved |
| use CRL representation instead of PointNav policy encoder | frozen encoder · ObjectNav (Matterport3D) | SPL: 0.0390 -> 0.0144 | Hurt |
| freeze convolution weights of a random network (ablation) | PointNav agent · PointNav | SPL | No clear effect |
| use CRL representation instead of From Scratch | Seq2Seq behavioral cloning · VLN val-unseen | SPL: 0.138 -> 0.157 | Improved |
| use CRL representation instead of ImageNet | Seq2Seq Dagger · VLN val-unseen | SPL: 0.206 -> 0.199 | Hurt |
| use CRL exploration instead of random exploration (ablation) | contrastive model · image diversity | LPIPS diversity: 0.708 -> 0.728 | Improved |
Closest papers by meaning
- An Empirical Investigation of Representation Learning for Imitation (NeurIPS 2021 · similarity 0.88)
- Causal Curiosity: RL Agents Discovering Self-supervised Experiments for Causal Representation Learning (ICML 2021 · similarity 0.86)
- Pre-trained Text-to-Image Diffusion Models Are Versatile Representation Learners for Control (NeurIPS 2024 · similarity 0.86)
- Selective Visual Representations Improve Convergence and Generalization for Embodied AI (ICLR 2024 · similarity 0.85)
- Towards Principled Representation Learning from Videos for Reinforcement Learning (ICLR 2024 · similarity 0.85)
- DMC-VB: A Benchmark for Representation Learning for Control with Visual Distractors (NeurIPS 2024 · similarity 0.84)
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.
