Research map / Reinforcement learning / Visual and representation RL
TGRL: An Algorithm for Teacher Guided 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 |
|---|---|---|---|
| Adding the teacher-guidance term with adaptive coefficient (TGRL) instead of pure imitation learning | student policy on Shadow Hand tactile pen reorientation · Shadow Hand pen reorientation (tactile) | success rate: 54% -> 73% | Helped |
| Using a sub-optimal teacher (0% success) vs perfect teacher for TGRL student (ablation) | TGRL student on Tiger Door · Tiger Door | student success rate: 100% with teachers at 100/80/40% but 0% with a 0% teacher | Mixed |
| Lowering teacher quality from 100% to 0% success (ablation) | TGRL student on Lava Crossing · Lava Crossing | student success rate: 100% -> 8% | Hurt |
| Joint replay buffer shared between main policy and auxiliary policy (vs separate buffers) (ablation) | TGRL on Light-Dark Ant · Light-Dark Ant | convergence rate | Helped |
| Adaptive balancing coefficient (dual update of lambda) instead of fixed coefficient (ablation) | TGRL on Shadow Hand · Shadow Hand pen reorientation (tactile) | asymptotic success | Helped |
| Using a fixed balancing coefficient (including the best-found value) instead of adaptive coefficient (ablation) | TGRL · fixed-coefficient sweep | final performance | Hurt |
| Adaptive coefficient vs constant alpha across initial values in Lava Crossing (ablation) | TGRL · Lava Crossing | convergence to optimal policy | Helped |
| Replacing TGRL with ADVISOR-off baseline | ADVISOR baseline · Tiger Door and Lava Crossing | success rate | Hurt |
| Using TGRL with mostly fixed hyperparameters (no per-task tuning) vs COSIL with 8-value hyperparameter search | COSIL baseline average across hyperparameters · Imitation Gap domains | success rate | Hurt |
| Adding TGRL on domains without Imitation Gap (proprioceptive Hopper, Walker2D, HalfCheetah with joint positions only) | TGRL vs IL · POMDP Mujoco without Imitation Gap | return vs teacher | No clear effect |
Closest papers by meaning
- Guarded Policy Optimization with Imperfect Online Demonstrations (ICLR 2023 · similarity 0.90)
- PEBBLE: Feedback-Efficient Interactive Reinforcement Learning via Relabeling Experience and Unsupervised Pre-training (ICML 2021 · similarity 0.87)
- DUO: Diverse, Uncertain, On-Policy Query Generation and Selection for Reinforcement Learning from Human Feedback (AAAI 2025 · similarity 0.87)
- Reinforcement Learning Guided Semi-Supervised Learning (NeurIPS 2024 · similarity 0.87)
- HYPRL: Reinforcement Learning of Control Policies for Hyperproperties (NeurIPS 2025 · similarity 0.86)
- Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents (NeurIPS 2026 · 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.