Research map / Reinforcement learning / Imitation learning
PIRLNav: Pretraining With Imitation and RL Finetuning for ObjectNav
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 BC pretraining on human demonstrations followed by PIRLNav RL fine-tuning rather than BC→RL-FT with the VPT fine-tuning approach | ObjectNav policy · HM3D val | Success: +2.2% Success | Helped |
| Add a critic-only learning phase before actor fine-tuning (ablation) | BC→RL-FT · HM3D val | Success: improves by 4.7% on success | Helped |
| Add critic learning-rate decay after critic learning (ablation) | BC→RL-FT · HM3D val | Success: improves success by 7.4% | Helped |
| Add actor learning-rate warmup after critic learning, without critic learning-rate decay (ablation) | BC→RL-FT · HM3D val | Success: improves success by 6.2% | Helped |
| Combine critic-only learning, critic learning-rate decay, and actor learning-rate warmup (ablation) | BC→RL-FT · HM3D val | Success: improves by 9.9% on success | Helped |
| Use human demonstrations rather than shortest-path demonstrations for BC pretraining | BC→RL ObjectNav policy · HM3D val | Success: 64.1% vs 6.4% | Helped |
| Use human demonstrations rather than frontier-exploration demonstrations for BC pretraining | BC→RL ObjectNav policy · HM3D val | Success: 64.1% vs 44.9% | Helped |
| Initialize RL fine-tuning from BC on human rather than frontier-exploration demonstrations, with BC train success controlled | ObjectNav policy · HM3D val | Success: 66.1% vs 51.3% | Helped |
| Initialize RL fine-tuning from BC on human rather than shortest-path demonstrations, with BC train success controlled | ObjectNav policy · HM3D val | Success: 66.1% vs 43.6% | Helped |
| Scale the human-demonstration BC dataset from 20k to 77k demonstrations | BC→RL ObjectNav policy · HM3D val | Success: BC→RL improvement is 10.1% ... at 20k ... 6.3% ... at 77k | Mixed |
| Use 35k rather than 77k human demonstrations for BC pretraining | BC→RL ObjectNav policy · HM3D val | Success: 66.4% vs 70.4% | Hurt |
| Increase frontier-exploration demonstrations beyond 70k | BC ObjectNav policy · HM3D val | Success | No clear effect |
Closest papers by meaning
- RLIF: Interactive Imitation Learning as Reinforcement Learning (ICLR 2024 · similarity 0.89)
- Dual RL: Unification and New Methods for Reinforcement and Imitation Learning (ICLR 2024 · similarity 0.89)
- Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning (NeurIPS 2024 · similarity 0.88)
- Limited Preference Aided Imitation Learning from Imperfect Demonstrations (ICML 2024 · similarity 0.88)
- Diffusion Model-Augmented Behavioral Cloning (ICML 2024 · similarity 0.88)
- Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning (NeurIPS 2025 · similarity 0.88)
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.