[Paper Review] Combining Optimal Control and Learning for Visual Navigation in Novel Environments

Somil Bansal, Varun Tolani+3 more|arXiv (Cornell University)|Mar 6, 2019

Robotic Path Planning AlgorithmsComputer Science96 citations

TL;DR

LB-WayPtNav pairs a learning-based perception module that predicts waypoints with a model-based planner and LQR-based tracking to navigate in unknown indoor environments, outperforming end-to-end learning and pure geometric mapping in simulation and real robot tests.

ABSTRACT

Model-based control is a popular paradigm for robot navigation because it can leverage a known dynamics model to efficiently plan robust robot trajectories. However, it is challenging to use model-based methods in settings where the environment is a priori unknown and can only be observed partially through on-board sensors on the robot. In this work, we address this short-coming by coupling model-based control with learning-based perception. The learning-based perception module produces a series of waypoints that guide the robot to the goal via a collision-free path. These waypoints are used by a model-based planner to generate a smooth and dynamically feasible trajectory that is executed on the physical system using feedback control. Our experiments in simulated real-world cluttered environments and on an actual ground vehicle demonstrate that the proposed approach can reach goal locations more reliably and efficiently in novel environments as compared to purely geometric mapping-based or end-to-end learning-based alternatives. Our approach does not rely on detailed explicit 3D maps of the environment, works well with low frame rates, and generalizes well from simulation to the real world. Videos describing our approach and experiments are available on the project website.

Motivation & Objective

• Motivate autonomous navigation in unknown, cluttered indoor environments without explicit 3D maps.
• Propose a hybrid framework that uses learning to predict collision-free waypoints and model-based control to generate feasible trajectories.
• Demonstrate that the approach generalizes from simulation to real robots and outperforms purely learning-based or mapping-based baselines.

Proposed method

• Perception module: a CNN predicts the next waypoint from onboard RGB image I_t, current speed u_t, and target p_t* in robot coordinates.
• Planning: fit third-order splines to generate smooth x,y trajectories from current state to the predicted waypoint, ensuring dynamic feasibility.
• Tracking: an LQR controller linearizes dynamics around the spline trajectory to produce feed-forward and feedback terms for execution over horizon H.
• Training: supervise perception using optimal waypoints derived from known maps during training, enabling zero-shot test-time performance without real-world labeling.
• Evaluation: compare against End-to-End learning and geometric mapping approaches in simulation and on a TurtleBot 2 hardware platform.

Experimental results

Research questions

• RQ1 Can a perception-driven waypoint predictor combined with model-based planning achieve reliable navigation in novel, cluttered indoor environments?
• RQ2 How does hybrid LB-WayPtNav compare to end-to-end learning and to mapping-based planning in terms of success rate, speed, and trajectory smoothness?
• RQ3 Does the approach generalize from simulation to real hardware without real-world finetuning?
• RQ4 What is the impact of using spline-based planning and LQR tracking on control smoothness and robustness to disturbances?

Key findings

• LB-WayPtNav achieves higher success rates and faster goal arrival than End-to-End learning in simulation and hardware tests.
• The method produces smoother trajectories with lower acceleration and jerk compared to pure end-to-end policies.
• In simulation, LB-WayPtNav outperforms a memoryless depth-based mapping baseline and approaches depth-based mapping performance when perfect depth is available.
• The approach generalizes from simulation to real robots with minimal domain gap and does not require explicit 3D maps during testing.
• Perception training uses optimal waypoints derived from known maps, enabling supervision without human labeling.