Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning Invariant Feature Spaces to Transfer Skills with 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 invariant feature-space transfer learned from shared proxy skills rather than learning from scratch | 4-link arm · Button pressing | success rate | Helped |
| Use learned invariant feature-space transfer rather than learning from scratch | Tendon-driven arm · Block pulling | task performance | Helped |
| Use learned invariant feature-space transfer from raw images rather than learning from scratch | 4-link arm · Block pushing | task performance | Helped |
| Use deep neural-network embeddings instead of CCA | Invariant feature-space transfer · Button pressing | learning speed | Helped |
| Use a direct source-to-target state mapping instead of a shared feature space | Invariant feature-space transfer · Button pressing | task performance | Hurt |
| Use KCCA with a quadratic kernel instead of the learned neural embedding | Transfer method · Button pressing | task performance | Hurt |
| Use KCCA with a quadratic kernel instead of the learned neural embedding | Transfer method · Block pulling | task performance | No clear effect |
| Learn embeddings from all three proxy tasks instead of one proxy task (ablation) | Invariant feature-space transfer · Button pressing | success rate | Helped |
| Use peg insertion as the sole proxy task instead of all three proxy tasks (ablation) | Invariant feature-space transfer · Button pressing | success rate | Hurt |
| Use block moving as the sole proxy task instead of all three proxy tasks (ablation) | Invariant feature-space transfer · Button pressing | success rate | Hurt |
| Use target reaching as the sole proxy task instead of all three proxy tasks (ablation) | Invariant feature-space transfer · Button pressing | success rate | Hurt |
| Estimate correspondences with the EM-style DTW procedure instead of time-based alignment (ablation) | Invariant feature-space transfer · Block pulling | task performance | No clear effect |
Closest papers by meaning
- Learning an Embedding Space for Transferable Robot Skills (ICLR 2018 · similarity 0.89)
- AutoCGP: Closed-Loop Concept-Guided Policies from Unlabeled Demonstrations (ICLR 2025 · similarity 0.88)
- CausalWorld: A Robotic Manipulation Benchmark for Causal Structure and Transfer Learning (ICLR 2021 · similarity 0.88)
- Learning transferable motor skills with hierarchical latent mixture policies (ICLR 2022 · similarity 0.88)
- Hierarchically Decoupled Imitation For Morphological Transfer (ICML 2020 · similarity 0.88)
- DittoGym: Learning to Control Soft Shape-Shifting Robots (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.