Research map / Reinforcement learning / Visual and representation RL
Task Tokens: A Flexible Approach to Adapting Behavior Foundation Models
Ron Vainshtein, Zohar Rimon, Shie Mannor, Chen Tessler
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 |
|---|---|---|---|
| adapt MaskedMimic with Task Tokens instead of full fine-tuning | MaskedMimic · Long Jump | success rate: 47.36 -> 99.75 | Improved |
| Task Tokens vs MaskedMimic fine-tuning on Strike | MaskedMimic · Strike | success rate: 83.07 vs 76.61 | Hurt |
| Task Tokens vs fine-tuning on Reach/Direction/Steering | MaskedMimic · Reach/Direction/Steering | success rate: 94.88 vs 93.70; 99.26 vs 99.10; 88.69 vs 87.44 | No clear effect |
| add Task Tokens over zero-shot MaskedMimic J.C. only | MaskedMimic · Reach | success rate: 24.77 -> 94.88 | Improved |
| use Task Tokens instead of PULSE on Steering | PULSE · Steering | success rate: 40.72 -> 88.69 | Improved |
| add joint conditioning to Task Tokens (ablation) | Task Tokens · Steering | success rate: 83.66 -> 88.69 | Improved |
| add joint conditioning to Task Tokens (ablation) | Task Tokens · Direction | success rate: 96.89 -> 99.26 | Improved |
| add joint conditioning to Task Tokens (ablation) | Task Tokens · Reach | success rate: 95.37 -> 94.88 | No clear effect |
| add joint conditioning to fine-tuning (ablation) | MaskedMimic F.T. · Steering | success rate: 87.44 -> 96.41 | Improved |
| use current pose input with smaller encoder, no J.C. (ablation) | Task Encoder · Steering | success rate: 84.28 -> 66.31 | Hurt |
| use bigger MLP encoder with J.C. and current pose (ablation) | Task Encoder · Steering | success rate: 79.47 -> 87.77 | Improved |
| Task Tokens vs MaskedMimic fine-tuning on human-likeness | MaskedMimic · human study, Direction | win rate: 99% ± 1% | Improved |
Closest papers by meaning
- Self-Improving Embodied Foundation Models (NeurIPS 2025 · similarity 0.89)
- QueST: Self-Supervised Skill Abstractions for Learning Continuous Control (NeurIPS 2024 · similarity 0.88)
- SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning Framework (NeurIPS 2026 · similarity 0.87)
- TokenHSI: Unified Synthesis of Physical Human-Scene Interactions through Task Tokenization (CVPR 2025 · similarity 0.87)
- Humanoid Locomotion as Next Token Prediction (NeurIPS 2024 · similarity 0.87)
- PRISE: LLM-Style Sequence Compression for Learning Temporal Action Abstractions in Control (ICML 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.
