Research map / Reinforcement learning / Visual and representation RL
Skew-Fit: State-Covering Self-Supervised 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 Skew-Fit goal sampling instead of standard RIG goal sampling | RIG · Visual Door, Visual Pusher, and Visual Pickup | goal-reaching performance and sample complexity | Helped |
| Use Skew-Fit goal sampling instead of the compared goal-sampling methods | Ant labyrinth · Ant labyrinth | exploration progress | Helped |
| Use Skew-Fit with RIG instead of standard RIG | Real World Visual Door · Real World Visual Door | success rate | Helped |
| Train the VAE with Skew-Fit rather than MLE on the unbalanced image dataset (ablation) | VAE · Visual Door uniform evaluation images | NLL: 20178.03 -> 20175.9 | Helped |
| Use a uniform image dataset rather than an unbalanced dataset (ablation) | VAE trained with MLE · Visual Door uniform evaluation images | NLL: 20178.03 -> 20175.4 | Helped |
| Set alpha to -1 rather than 0 (no Skew-Fit) (ablation) | Skew-Fit · Visual Door and Visual Pusher | task performance | Helped |
| Replace SAC with TD3 while retaining Skew-Fit (ablation) | Skew-Fit · Visual Door, Visual Pusher, and Visual Pickup | task performance | No clear effect |
| Use SIR rather than importance sampling to train the generative model (ablation) | Skew-Fit · Visual Door unbalanced-image variance experiment | gradient variance | Helped |
| Use importance sampling rather than SIR to train the generative model (ablation) | Skew-Fit · Visual Door unbalanced-image variance experiment | gradient variance | Hurt |
| Initially train the VAE with standard MLE before applying Skew-Fit (ablation) | Skew-Fit · Visual continuous-control tasks | training stability | Helped |
| Use Skew-Fit with alpha values across the tested range rather than alpha=0 (ablation) | Skew-Fit · Visual Door and Visual Pusher | task performance | Helped |
Closest papers by meaning
- Learning to Reach Goals via Iterated Supervised Learning (ICLR 2021 · similarity 0.88)
- Model-Based Visual Planning with Self-Supervised Functional Distances (ICLR 2021 · similarity 0.88)
- C-Planning: An Automatic Curriculum for Learning Goal-Reaching Tasks (ICLR 2022 · similarity 0.87)
- Exploring the Edges of Latent State Clusters for Goal-Conditioned Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.86)
- Stabilizing Contrastive RL: Techniques for Robotic Goal Reaching from Offline Data (ICLR 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.