Active Visual SLAM with Independently Rotating Camera

Elia Bonetto, Pascal Goldschmid, Michael J. Black, Aamir Ahmad

TL;DR

This paper introduces a novel approach to Active Visual SLAM (V-SLAM) by equipping a robot with an independently rotating camera. This allows for more efficient mapping and localization by optimizing camera viewpoints without necessarily moving the entire robot, leading to reduced energy consumption and improved map accuracy, even on non-omnidirectional platforms.

What Problem Does This Paper Address?

Traditional Visual SLAM (V-SLAM) systems often rely on cameras rigidly attached to a robot's body. While active V-SLAM aims to improve mapping by intelligently controlling the robot's movement to find optimal viewpoints, this fixed camera configuration limits the robot's maneuverability and the potential for fine-tuning visual information acquisition. This limitation means that the robot must often rotate its entire body to achieve a desired camera angle, which is energy-intensive and can be less flexible. Furthermore, existing active SLAM methods often assume perfect calibration or ignore the uncertainties introduced by additional degrees of freedom, like an independently moving camera. This paper tackles these issues by proposing a system that leverages an independently rotating camera and explicitly accounts for the associated estimation uncertainties.

Methods

The core innovation is the introduction of an independently rotating camera on the robot's base. This provides an additional degree of freedom for controlling the camera's heading, allowing for optimal viewpoint selection without rotating the entire robot. To handle the increased complexity and potential estimation uncertainties, the authors propose a novel state estimation formulation that merges the robot's pose and the camera's relative orientation into a unified state. This merged state is then fed into an Extended Kalman Filter (EKF) for sensor fusion, incorporating data from IMUs and wheel odometry. The control aspect utilizes an NMPC (Nonlinear Model Predictive Control) formulation, adapted to account for the combined robot and camera state. The method is built upon a state-of-the-art active V-SLAM approach for omnidirectional robots but is designed to generalize to other robotic platforms.

Datasets

The paper evaluates its method using two simulated environments: "AWS Robomaker Small House World" and a modified "Gazebo's Cafè environment." For real-world experiments, the authors use their physical robot platform (Festo Didactic's Robotino) in an office environment. The ground truth maps for simulation were generated using pgm_map_creator.

Key Insights and Ideas

  • Independent Camera Rotation is Key: Adding an extra degree of freedom through an independently rotating camera significantly enhances active SLAM capabilities, enabling more efficient exploration and mapping.
  • Uncertainty Matters: Ignoring the uncertainties introduced by an independently moving camera can negatively impact SLAM performance. Merging the robot and camera state estimates and jointly estimating their uncertainties is crucial for robust performance.
  • Generalization is Possible: The proposed method is not limited to omnidirectional robots, demonstrating its applicability to other robotic platforms, which was a limitation of previous approaches.
  • Energy Efficiency Gains: By allowing the camera to rotate independently, the system can achieve lower overall wheel rotation, leading to reduced energy consumption.
  • Semi-holonomic Control Benefits: In certain scenarios, like the semi-holonomic robot (where only translation is allowed), the independent camera rotation proves particularly effective, leading to better performance metrics.

Key Results

  • Improved Map Accuracy: The proposed method consistently achieves better Balanced Accuracy (BAC) scores compared to baselines, with improvements ranging from ~1.5% to ~5%.
  • Reduced Trajectory Error: Absolute Trajectory Error (ATE) RMSE is reduced by ~10% to ~40% when using the merged state estimate.
  • More Frequent Loop Closures: The number of loop closures per meter traveled increases by ~10% to ~30% in most cases, indicating better place recognition.
  • Lower Energy Consumption: The system demonstrates a reduction in total wheel rotation per meter traveled, suggesting significant energy savings.
  • Comparable Mapping Performance: While the active SLAM approach itself influences exploration, the core mapping performance (e.g., map entropy) remains comparable to baselines when the merged state estimate is used.
  • Real-world Validation: Experiments on a physical robot confirm the benefits observed in simulation, showing improved exploration speed and shorter paths.

Limitations and Future Work

  • PID Delay in Real-world: The real-world experiments were affected by a PID controller's delay in managing camera rotation, which impacted performance. Future work could explore more advanced control strategies.
  • Encoder Limitations: The absence of a camera encoder in the real robot experiments made it challenging to directly distinguish between robot and camera rotations, relying on approximations.
  • Non-holonomic Optimization: The authors acknowledge that optimizing for non-holonomic robots was not fully explored, suggesting this as a future direction.
  • Further Exploration of Control: Investigating more sophisticated control strategies for different robotic platforms (e.g., full holonomic control for energy optimization) is suggested.
  • Hardware Improvements: The paper hints at potential hardware improvements that could further enhance performance.

Summary updated

Abstract

In active Visual-SLAM (V-SLAM), a robot relies on the information retrieved by its cameras to control its own movements for autonomous mapping of the environment. Cameras are usually statically linked to the robot’s body, limiting the extra degrees of freedom for visual information acquisition. In this work, we overcome the aforementioned problem by introducing and leveraging an independently rotating camera on the robot base. This enables us to continuously control the heading of the camera, obtaining the desired optimal orientation for active V-SLAM, without rotating the robot itself. However, this additional degree of freedom introduces additional estimation uncertainties, which need to be accounted for. We do this by extending our robot’s state estimate to include the camera state and jointly estimate the uncertainties. We develop our method based on a state-of-the-art active V-SLAM approach for omnidirectional robots and evaluate it through rigorous simulation and real robot experiments. We obtain more accurate maps, with lower energy consumption, while maintaining the benefits of the active approach with respect to the baseline. We also demonstrate how our method easily generalizes to other non-omnidirectional robotic platforms, which was a limitation of the previous approach. Code and implementation details are provided as open-source.
BibTeX
@inproceedings{bonetto2021active,
  author = {Elia Bonetto and Pascal Goldschmid and Michael J. Black and Aamir Ahmad},
  title = {{Active Visual SLAM with Independently Rotating Camera}},
  booktitle = {ECMR},
  year = {2021},
  url = {https://openreview.net/forum?id=oXR_QOgtBBm},
}
Figure