Research map / Reinforcement learning / Visual and representation RL
Reinforcement Learning via Auxillary Task Distillation
Abhinav N Harish, Larry Heck, Josiah P Hanna, Zsolt Kira, Andrew Szot
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 |
|---|---|---|---|
| add the distillation loss from auxiliary tasks (ablation) | AuxDistill · Habitat Rearrangement (Eval Unseen, all episodes) | success rate (%): 0 -> 52 | Improved |
| train auxiliary and main tasks concurrently with distillation instead of a two-stage RL curriculum (ablation) | AuxDistill · Habitat Rearrangement (Eval Unseen, all episodes) | success rate (%): 0 -> 52 | Improved |
| include the Pick auxiliary task (ablation) | AuxDistill · Rearrangement fridge subset, Eval | success rate (%): 0 -> 30 | Improved |
| add the Pick from Fridge auxiliary task (ablation) | AuxDistill · Rearrangement fridge subset, Eval hard | success rate (%): 7 -> 19 | Improved |
| add the Open Fridge auxiliary task (ablation) | AuxDistill · Rearrangement fridge subset, Eval all episodes | success rate (%): 48 -> 30 | Hurt |
| increase distillation coefficient lambda from 0.1 to 1.0 (ablation) | AuxDistill · Rearrangement (Eval all episodes) | success rate (%): 30 -> 1 | Hurt |
| reduce distillation coefficient lambda to 0.01 (ablation) | AuxDistill · Rearrangement (Eval all episodes) | success rate (%): 36 -> 10 | Hurt |
| use coordinate-pick auxiliary task with distillation for Category Pick | AuxDistill · Category Pick (Eval Unseen) | success rate (%): 9 -> 11 | Improved |
| replace pre-trained skill hierarchy (M3) with end-to-end AuxDistill | Habitat rearrangement · Eval Unseen Hard | success rate (%): 0 -> 41 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Galactic — Galactic: Scaling End-to-End Reinforcement Learning for Rearrangement at 100k Steps-per-Second (CVPR 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents (NeurIPS 2026 · similarity 0.87)
- Multi-Stage Manipulation with Demonstration-Augmented Reward, Policy, and World Model Learning (ICML 2025 · similarity 0.87)
- DiffTORI: Differentiable Trajectory Optimization for Deep Reinforcement and Imitation Learning (NeurIPS 2024 · similarity 0.87)
- Robot-R1: Reinforcement Learning for Enhanced Embodied Reasoning in Robotics (NeurIPS 2025 · similarity 0.87)
- Policy Compatible Skill Incremental Learning via Lazy Learning Interface (NeurIPS 2025 · similarity 0.87)
- PEAR: Primitive Enabled Adaptive Relabeling for Boosting Hierarchical Reinforcement Learning (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.
