Abstract:In autonomous robot navigation, terrain cost assignment is typically performed using a semantics-based paradigm in which terrain is first labeled using a pre-trained semantic classifier and costs are then assigned according to a user-defined mapping between label and cost. While this approach is rapidly adaptable to changing user preferences, only preferences over the types of terrain that are already known by the semantic classifier can be expressed. In this paper, we hypothesize that a machine-learning-based alternative to the semantics-based paradigm above will allow for rapid cost assignment adaptation to preferences expressed over new terrains at deployment time without the need for additional training. To investigate this hypothesis, we introduce and study PACER, a novel approach to costmap generation that accepts as input a single birds-eye view (BEV) image of the surrounding area along with a user-specified preference context and generates a corresponding BEV costmap that aligns with the preference context. Using both real and synthetic data along with a combination of proposed training tasks, we find that PACER is able to adapt quickly to new user preferences while also exhibiting better generalization to novel terrains compared to both semantics-based and representation-learning approaches.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: How can robots quickly adapt to new terrain preferences during the navigation process and generate costmaps consistent with these preferences, thereby achieving path planning that is more in line with human expectations? Specifically, although existing semantic - based methods can quickly adapt to preferences for known terrain types, they cannot handle unknown terrain types; while representation - learning - based methods can theoretically be generalized to new terrains, in practice, the utility function needs to be retrained to adapt to new preference rankings, which makes them less flexible for changing operator preferences.
To solve these problems, the paper proposes a new method named PACER (Preference - conditioned All - terrain Costmap Generation). PACER generates a costmap consistent with these preferences by taking a bird - eye - view (BEV image) and user - specified preference context as input. This method can not only quickly adapt to new user preferences during deployment but also better generalize to previously unseen terrain types.
### Core Problems and Solutions of PACER
1. **Quick Adaptation to New Preferences**:
- Existing methods (such as semantic - based methods) can only handle preferences for known terrain types, while PACER can quickly adjust the costmap according to the new preference context during deployment.
2. **Generalization to Unknown Terrains**:
- Although representation - learning - based methods can theoretically be generalized to new terrains, in practice, the utility function needs to be retrained. PACER can handle unknown terrains without retraining by using neural network architectures and specific training strategies.
3. **Generation of Costmaps Consistent with Human Preferences**:
- PACER ensures that the generated costmap is consistent with the preference context provided by the user, making the robot's navigation behavior more in line with human expectations.
### Main Contributions of the Paper
- **Proposing the PACER Model**: A novel costmap generation method that can quickly adapt to new user preferences during deployment and generalize to unknown terrains.
- **Designing Three Training Phases**: Training for real - world data, changed preferences, and synthetic terrains respectively to ensure the robustness and generalization ability of the model in different scenarios.
- **Experimental Verification**: Proving the effectiveness and superiority of PACER on known and unknown terrains through experiments, especially showing better generalization ability when dealing with unseen terrains.
### Formula Summary
In the paper, some of the key formulas involved are as follows:
- **Total Objective Function for Path Planning**:
\[
\Gamma_S=\arg\min_\Gamma \|x_S - G\|+\lambda J(\Gamma)
\]
where \(\|x_S - G\|\) is the distance between the final state \(x_S\) and the target location \(G\), \(J(\Gamma)\) is the cost function of the path, and \(\lambda\) is the relative weight.
- **Terrain Cost Function**:
\[
J(\Gamma)=J_{\text{geometric}}(\Gamma)+\alpha J_{\text{social}}(\Gamma)+\beta J_{\text{terrain}}(\Gamma)
\]
where \(J_{\text{geometric}}(\Gamma)\) is the geometric cost, \(J_{\text{social}}(\Gamma)\) is the social navigation cost, \(J_{\text{terrain}}(\Gamma)\) is the terrain cost, and \(\alpha\) and \(\beta\) are relative weights.
- **Loss Function**:
\[
\phi^*=\arg\min_\phi \mathbb{E}_{(I, \hat{H}, C_T)\sim D} \left[ \ell(\hat{R}_\phi(I, \hat{H}), C_T) \right]