A project for social navigation in embodied environments.
This repository is built on top of previous work, and we further adapted, trained, and evaluated the system for our own experiments.
SocialNav2 focuses on PointGoal social navigation in 3D indoor environments, where an agent must:
- 🎯 reach the target position,
- 🚶 avoid dynamic human obstacles,
- 🧠 navigate efficiently,
- 🤝 maintain social compliance.
Our current approach uses a two-stage training paradigm:
- Imitation Learning (IL) for policy pre-training
- Reinforcement Learning (RL) for fine-tuning
Our recent progress is based on a two-stage pipeline:
-
📦 Expert Data Collection
We first collect high-quality expert trajectories in Habitat and filter them to keep only successful, high-reward episodes. -
🧠 Imitation Learning Pre-training
We use Behavioral Cloning (BC) to train the policy from expert demonstrations, giving the agent a strong initial navigation behavior. -
🔥 Reinforcement Learning Fine-tuning
We then fine-tune the pretrained model with PPO, allowing the agent to further improve success rate, path efficiency, and collision avoidance. -
👀 Social Awareness Enhancement
During RL fine-tuning, we incorporate auxiliary socially-aware objectives such as:- human count estimation,
- current human position tracking,
- future trajectory forecasting.
Our policy network includes:
- 🖼️ a visual encoder for egocentric observations,
- 📍 a state encoder for pose and goal information,
- 🧠 a GRU-based temporal memory module,
- 🎯 a policy head and a value head for action prediction and RL optimization.
Our current experiments show that the IL + RL pipeline performs better than both:
- training with RL only from scratch, and
- using IL only without RL fine-tuning.
| Method | SR ↑ | SPL ↑ | PSC ↑ | H-Coll ↓ | Total ↑ |
|---|---|---|---|---|---|
| Baseline (RL-only) | 0.5400 | 0.4997 | 0.8630 | 0.3920 | 0.6248 |
| Ours (IL-only) | 0.6280 | 0.5761 | 0.8609 | 0.3540 | 0.6823 |
| Ours (IL+RL) | 0.6480 | 0.6010 | 0.8607 | 0.3420 | 0.6977 |
These results show that:
- ✅ IL provides a strong initialization,
- ✅ RL fine-tuning improves efficiency and robustness,
- ✅ the combined method achieves the best overall performance.
- Habitat simulator
- HM3D environments
- Bullet physics engine
This repository is based on previous social navigation work.
We built on top of the existing project and further adapted it for our own training, experiments, and evaluation.
Current focus:
- improving social compliance,
- reducing human collision rate,
- building stronger embodied navigation policies,
- and exploring better training strategies for dynamic environments.