Learning the input space of the ego-agent for Social Navigation using cost-guided optimization
Published:
My work for IROS 2025 is to improve CrowdSurfer and integrate social behavior into it. This has been done by training a Vision Transformer(ViT) based classifier that maps the observation space of CrowdSurfer to a set of socially-compliant rules. The observation space includes:
- Occupancy maps
- Dynamic obstacle positions
- Static obstacle positions
The classifier outputs a softmax distribution over these rules, which are then mapped to a set of hand-tuned cost functions. Below is an example of one such cost function, Group Interaction Cost:
\[C_{group}(x, y) = \sum_{j=1}^{M} \frac{1}{d_j + \epsilon}\]- M: Number of groups detected in the environment
- dj: Distance to the centroid of the j-th group
- ε: A small positive constant to avoid division by zero when dj is very small
This cost ensures the ego-agent avoids passing through groups of people.
Additional Cost Functions
1. Directional Alignment Cost
\[C_{alignment}(x, y) = \sum_{t} \left( 1 - \cos\bigl(\theta_{agent,t} - \theta_{flow,t}\bigr) \right)\]Encourages socially compliant behavior by matching the direction of nearby pedestrians.
2. Yielding Cost
\[C_{yielding} = w_y \cdot \sum_{j=1}^{m} \frac{\bigl(1 - P_{\text{priority}}(j)\bigr)}{d_j + \epsilon} \cdot \max\bigl(0, v_{j,\text{rel}} - v_{\text{threshold}}\bigr)\]Proximity Penalty \(\frac{1}{d_j + \epsilon}\):
Closer obstacles contribute more to the cost, encouraging the agent to yield when obstacles are nearby.Relative Velocity Contribution \(\max(0, v_{j,\text{rel}} - v_{\text{threshold}})\):
The cost increases only if the relative velocity exceeds a predefined threshold \(v_{\text{threshold}}\). This ensures the agent responds to dynamic obstacles with significant motion differences.Priority Weighting \(\bigl(1 - P_{\text{priority}}(j)\bigr)\):
Obstacles with higher priority contribute less to the cost, encouraging the agent to yield to them.
Features
- Current Implementation: Includes only Group Interaction Cost.
- ViT Training: Conducted on a synthetic dataset of PEDSIM rosbags.
Technologies Used
- Optimal Control
- Machine Learning
- PyTorch
- JAX
Note: The project repository is still under development and remains private.