Abstract
This paper presents a novel approach to active mapping where a robot autonomously optimizes its trajectory to maximize information gain. By developing a differentiable approximation of Shannon Mutual Information (SMI), the authors enable the use of continuous gradient-based optimization on the $SE(3)$ manifold. This method allows robots to "nudge" their paths toward high-information areas while simultaneously accounting for obstacle avoidance and view redundancy, overcoming the limitations of traditional discrete candidate-based search methods.
Introduction
When a robot explores an unknown environment, it must plan paths that maximize the information gathered to build an accurate map. This process, known as active mapping, typically relies on maximizing Shannon Mutual Information (SMI) between the current map state and future sensor measurements. However, calculating SMI traditionally involves ray-tracing through discrete grid-based maps, resulting in a piecewise constant function. Because this function lacks meaningful derivatives, standard calculus-based optimization (gradients) cannot be used. Most existing systems are forced to sample a discrete set of candidate poses, which is computationally expensive and often misses the true global optimum.
Methodology
The core contribution of this work is a differentiable approximation of SMI that operates over the continuous space of 3D poses ($SE(3)$). The researchers achieve this through several key innovations:
- Differentiable Interpolation: Instead of direct calculation at every pose, the system defines a "viewpoint grid." The SMI at any specific location is calculated as a weighted average of the SMI at nearby grid points using a smooth weighting function, ensuring the resulting SMI surface is differentiable.
- Gradient Ascent on $SE(3)$: With a smooth function, the robot can use gradient ascent to find the direction of maximum information gain. The authors utilize Lie Group theory to ensure that the robot's orientation remains mathematically valid throughout the optimization process.
- Multi-Objective Optimization: The objective function incorporates differentiable penalty terms for collisions (using distance transforms) and redundancy (discouraging overlapping fields of view).
Differentiable SMI Approximation
The smoothing of the "information field" is the critical breakthrough. By interpolating between pre-calculated points, the robot can "slide" along the SMI surface to find peaks. To maintain real-time performance, the authors pre-compute ray-tracings from a single origin point and translate them to cell centers, avoiding redundant geometric calculations.
As shown in the comparison below, the exact SMI function is highly irregular and non-differentiable. The proposed approximation ($\xi_{max}$) smooths these transitions, providing a reliable gradient for optimization while maintaining the general profile of the information landscape.
Experimental Evaluation and Results
The method was rigorously tested across simulated 2D environments, a photo-realistic 3D Unity simulation with a Husky robot, and real-world indoor experiments using an NVIDIA Xavier NX-powered wheeled robot.
- Exploration Efficiency: In 2D trials, the gradient-based method explored environments faster and reduced map entropy more effectively than standard frontier-based or sequential SMI baselines.
- Safety and Clearance: The differentiable collision penalty significantly improved safety. In 3D simulations, the robot maintained an average obstacle clearance of 3.4–3.9m, a substantial improvement over the 1.2m clearance achieved by frontier-based methods.
- Real-Time Performance: The system proved viable for embedded hardware, updating 3D OctoMaps every 0.11 seconds.
- Dimensionality Trade-offs: While 2D projections (Exp-2D) offered faster planning, full 3D planning (Exp-3D) was more efficient in terms of total distance traveled, as it better accounted for complex 3D occlusions.
Conclusion
By moving from discrete search to continuous optimization, this framework allows robots to find more informative sensing views that discrete samplers would likely overlook. While the method can occasionally encounter local optima—a common challenge in gradient ascent—initializing the path with a frontier-based planner effectively mitigates this risk. Future work aims to scale this differentiable framework to multi-agent systems, where robots can coordinate to minimize redundant exploration.
Summary updated