Research map / Reinforcement learning / Visual and representation RL
Dynamical Distance Learning for Semi-Supervised and Unsupervised Skill Discovery
Kristian Hartikainen, Xinyang Geng, Tuomas Haarnoja, Sergey Levine
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 supervised regression on on-policy trajectory pairs to learn dynamical distances | DDL · DClaw valve turning (state and vision) | task success | Improved |
| Use cumulative distance as the policy objective instead of greedily minimizing distance (ablation) | DDL · DClaw valve turning (state and vision) | task learning | Improved |
| Replace supervised distance regression with temporal-difference Q-learning (ablation) | DDL · DClaw valve turning (state and vision) | task learning | Hurt |
| Use hindsight experience replay with preference goals instead of dynamical-distance rewards (ablation) | DDLfP · DClaw valve turning (state and vision) | task learning | Mixed |
| Use dynamical-distance rewards rather than a per-step reward of -1 until the goal | Goal-reaching policy learning · DClaw valve turning (state and vision) | task learning | Improved |
| Use 100 preference queries with DDLfP rather than the prior preference-learning method's 750 queries | DDLfP · MuJoCo locomotion tasks | true task return | Mixed |
| Select goals at the greatest estimated dynamical distance from the initial state | DDLUS · Hopper-v3, HalfCheetah-v3, Ant-v3 | distance traveled | Improved |
| Select goals at the greatest estimated dynamical distance from the initial state | DDLUS · InvertedDoublePendulum-v2 | task performance | Improved |
| Use preference queries to specify goals instead of manually designed rewards or goal examples | DDLfP · Real-world DClaw valve turning | valve rotation | Improved |
Closest papers by meaning
- Model-Based Visual Planning with Self-Supervised Functional Distances (ICLR 2021 · similarity 0.91)
- Generalizing Skills with Semi-Supervised Reinforcement Learning (ICLR 2017 · similarity 0.89)
- TimeRewarder: Learning Dense Reward from Passive Videos via Frame-wise Temporal Distance (ICML 2026 · similarity 0.88)
- Actionable Models: Unsupervised Offline Reinforcement Learning of Robotic Skills (ICML 2021 · similarity 0.88)
- Periodic Skill Discovery (NeurIPS 2025 · similarity 0.87)
- METRA: Scalable Unsupervised RL with Metric-Aware Abstraction (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.
