Research map / Reinforcement learning / Visual and representation RL
Weakly-Supervised Reinforcement Learning for Controllable Behavior
Lisa Lee, Ben Eysenbach, Ruslan Salakhutdinov, Shixiang (Shane) Gu, Chelsea Finn
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 |
|---|---|---|---|
| Use WSC goal generation in disentangled latent space instead of VAE-space goals | SkewFit · Sawyer visual goal tasks | success vs steps | Improved |
| Add auxiliary factor-identity prediction loss to SkewFit (Skewfit+pred) (ablation) | SkewFit · Sawyer visual goal tasks | success vs steps | Hurt |
| Use disentangled distance for reward only, goals sampled in VAE space (SkewFit+DR) (ablation) | SkewFit · Push n=3 | success vs steps | Mixed |
| Replace WSC goal generation by VAE-space goals (ablation) | WSC · Sawyer tasks | success vs steps | Hurt |
| Increase number of weak labels N from 128 to 4096 (ablation) | WSC disentangled model · PushLights n=3 obj1_x | Pearson correlation: 0.44 -> 0.93 | Improved |
| Use 1024 pairwise labels (ablation) | WSC disentangled model · all domains | Pearson correlation: 0.8 or higher | Improved |
| Increase number of weak labels (ablation) | WSC · visual goal-conditioned RL | task performance: 1024 labels sufficient | Improved |
| WSC with only 128 labels vs SkewFit on DoorLights (ablation) | WSC · DoorLights | task performance: 128 labels | Improved |
| Drop weak labels for task-irrelevant factors (hand position, lighting) (ablation) | WSC · PushLights | task performance | No clear effect |
| Use weak supervision instead of unsupervised VAE for disentanglement (ablation) | WSC · PushLights n=3 obj1_x | Pearson correlation: 0.02 -> 0.86 | Improved |
| Corrupt 5% of weak labels (ablation) | WSC · PushLights n=1 | Pearson correlation (All): 0.797 | Mixed |
| Raise label noise from 5% to 10% (ablation) | WSC · PushLights n=1 | Pearson correlation (All): 0.797 -> 0.638 | Hurt |
Closest papers by meaning
- Generalizing Skills with Semi-Supervised Reinforcement Learning (ICLR 2017 · similarity 0.87)
- Dynamical Distance Learning for Semi-Supervised and Unsupervised Skill Discovery (ICLR 2020 · similarity 0.87)
- Imagine Beyond ! Distributionally Robust Autoencoding for State Space Coverage in Online Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Prioritized Soft Q-Decomposition for Lexicographic Reinforcement Learning (ICLR 2024 · similarity 0.86)
- Excluding the Irrelevant: Focusing Reinforcement Learning through Continuous Action Masking (NeurIPS 2024 · similarity 0.86)
- Efficient Exploration and Discriminative World Model Learning with an Object-Centric Abstraction (ICLR 2025 · 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.
