Research map / Reinforcement learning / Visual and representation RL
DD-PPO: Learning Near-Perfect PointGoal Navigators from 2.5 Billion Frames
Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, Dhruv Batra
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 synchronous gradient all-reduce with rollout preemption for stragglers | PPO training in resource-intensive Habitat simulation · Homogeneous workload, 256 GPUs | speedup: 196x | Improved |
| preempt straggler rollouts at 80% or 60% completion threshold instead of waiting for all workers | DD-PPO · Heterogeneous Gibson and Matterport3D workload | scaling | Improved |
| train on Gibson-4+ plus all Matterport3D scenes instead of Gibson-4+ alone (ablation) | DD-PPO ResNet50 RGB-D · Habitat Challenge 2019 val | SPL: 0.922 ± 0.004 -> 0.956 ± 0.002 | Improved |
| add lower-quality Gibson-2/3 reconstructions to training (ablation) | DD-PPO ResNet50 RGB-D · Habitat Challenge 2019 val | SPL: 0.922 ± 0.004 -> 0.956 ± 0.003 | Improved |
| replace ResNet50 with SE-ResNeXt50 (ablation) | DD-PPO RGB-D trained on Gibson-2+ · Habitat Challenge 2019 val | SPL: 0.956 ± 0.003 -> 0.959 ± 0.002 | Improved |
| increase visual encoder capacity to SE-ResNeXt101 and LSTM hidden size to 1024-d (ablation) | DD-PPO RGB-D trained on Gibson-2+ · Habitat Challenge 2019 val | SPL: 0.959 ± 0.002 -> 0.969 ± 0.002 | Improved |
| train RGB navigation with Gibson-2+, all Matterport3D, and channel-wise RGB normalization | DD-PPO RGB agent · Habitat Challenge 2019 val | SPL: 0.57 -> 0.929 | Improved |
| remove GPS+Compass and train the RGB agent for 2.5 billion steps | DD-PPO navigation · PointGoal navigation without GPS+Compass | SPL: 0 -> 0.15 | Improved |
| initialize the visual encoder from PointGoalNav rather than ImageNet pre-training | Flee and Exploration transfer agents · Flee and Exploration | task performance | Improved |
| initialize and fine-tune the PointGoalNav policy for a new task | Flee and Exploration transfer agents · Flee and Exploration | learning speed | Improved |
Later papers that compared with this method
3 later accepted papers report DD-PPO as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- NavOL: Navigation Policy with Online Imitation Learning (ICML 2026 · better on 2 of 2 shared results)
- VER: Scaling On-Policy RL Leads to the Emergence of Navigation in Embodied Rearrangement (NeurIPS 2022 · better on 4 of 4 shared results)
- Learning Active Camera for Multi-Object Navigation (NeurIPS 2022 · better on 2 of 2 shared results)
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
- VER: Scaling On-Policy RL Leads to the Emergence of Navigation in Embodied Rearrangement (NeurIPS 2022 · similarity 0.89)
- Learning to Navigate Efficiently and Precisely in Real Environments (CVPR 2024 · similarity 0.86)
- GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach (CVPR 2025 · similarity 0.85)
- Diffusion Policy Policy Optimization (ICLR 2025 · similarity 0.85)
- Learn to Follow: Decentralized Lifelong Multi-Agent Pathfinding via Planning and Learning (AAAI 2024 · similarity 0.85)
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.
