[Paper Review] SLAM-Safe Planner: Preventing Monocular SLAM Failure using Reinforcement Learning.

Vignesh Prasad, Saurabh Singh+3 more|arXiv (Cornell University)|Jul 26, 2016

Robotics and Sensor-Based Localization19 references3 citations

TL;DR

This paper proposes a reinforcement learning (RL)-based planner, SLAM-Safe Planner, that generates trajectories to prevent monocular SLAM failure by learning safe camera motion policies. By training on perceptual inputs and motor actions, the RL agent produces trajectories that maintain SLAM robustness, significantly improving map quality and trajectory accuracy in simulations compared to baseline methods.

ABSTRACT

Effective SLAM using a single monocular camera is highly preferred due to its simplicity. However, when compared to trajectory planning methods using depth-based SLAM, Monocular SLAM in loop does need additional considerations. One main reason being that for the optimization, in the form of Bundle Adjustment (BA), to be robust, the SLAM system needs to scan the area for a reasonable duration. Most monocular SLAM systems do not tolerate large camera rotations between successive views and tend to breakdown. Other reasons for Monocular SLAM failure include ambiguities in decomposition of the Essential Matrix, feature-sparse scenes and more layers of non linear optimization apart from BA. This paper presents a novel formulation based on Reinforcement Learning (RL) that generates fail safe trajectories wherein the SLAM generated outputs (scene structure and camera motion) do not deviate largely from their true values. Quintessentially, the RL framework successfully learns the otherwise complex relation between motor actions and perceptual inputs that result in trajectories that do not cause failure of SLAM, which are almost intractable to capture in an obvious mathematical formulation. We show systematically in simulations how the quality of the SLAM map and trajectory dramatically improves when trajectories are computed by using RL.

Motivation & Objective

• To address the instability of monocular SLAM under large camera rotations and feature-sparse conditions.
• To reduce SLAM failure caused by ambiguous essential matrix decomposition and poor initialization in bundle adjustment.
• To develop a trajectory planner that ensures perceptual consistency and SLAM robustness without relying on explicit mathematical formulations of failure conditions.
• To improve the reliability of monocular SLAM in loop-closure scenarios through learned, safe motion policies.

Proposed method

• The approach formulates trajectory planning as a reinforcement learning problem to learn safe camera motion policies.
• The RL agent receives perceptual inputs (e.g., visual features, motion estimates) and selects actions (camera motion commands) to minimize SLAM error.
• The reward function is designed to penalize large deviations in SLAM outputs (structure and motion) from ground truth, promoting stability.
• The framework is trained in simulation using environments with challenging monocular SLAM conditions, including sparse features and large rotations.
• The policy is optimized using deep RL algorithms to map visual observations to actions that preserve SLAM integrity.
• The method avoids explicit modeling of SLAM failure modes by learning from failure signals in the reward function.

Experimental results

Research questions

• RQ1 Can a reinforcement learning agent learn to generate trajectories that prevent monocular SLAM failure in complex environments?
• RQ2 How does the RL-based planner compare to conventional trajectory planners in maintaining SLAM accuracy and robustness?
• RQ3 To what extent can RL learn the implicit relationship between camera motion and SLAM stability without explicit failure modeling?
• RQ4 Does the RL policy generalize across diverse scene structures and motion profiles in simulation?

Key findings

• The SLAM-Safe Planner significantly improves the quality of the reconstructed scene structure and camera trajectory in simulation.
• Trajectories generated by the RL planner result in lower SLAM error compared to baseline methods, especially under large rotation and feature-sparse conditions.
• The RL agent successfully learns to avoid camera motions that trigger SLAM failure, even without explicit modeling of failure mechanisms.
• The method demonstrates robustness to ambiguous essential matrix decomposition and poor bundle adjustment convergence by learning safe motion patterns.
• The improvement in map quality is quantitatively measurable, with reduced drift and more consistent structure estimates.