Differentiable Composite Neural Signed Distance Fields for Robot Navigation in Dynamic Indoor Environments

TL;DR

This work tackles dynamic indoor robot navigation using only onboard RGB-D sensing by proposing a compositional neural SDF framework that separates object-level and scene-level representations. A two-branch trajectory optimization strategy—Robot Body SDF as the fast primary path and Scene SDF as a slower, more informed fallback—achieves high success rates with reduced amortized planning time. The approach uses memory modules to handle limited field-of-view and a dual-mode mechanism to avoid local minima, validated in iGibson 2.0 and in real-world Turtlebot4 experiments. The results demonstrate improved reliability and flexibility for collision-aware planning in cluttered, dynamic environments, with modularity to substitute components as better models become available.

Abstract

Neural Signed Distance Fields (SDFs) provide a differentiable environment representation to readily obtain collision checks and well-defined gradients for robot navigation tasks. However, updating neural SDFs as the scene evolves entails re-training, which is tedious, time consuming, and inefficient, making it unsuitable for robot navigation with limited field-of-view in dynamic environments. Towards this objective, we propose a compositional framework of neural SDFs to solve robot navigation in indoor environments using only an onboard RGB-D sensor. Our framework embodies a dual mode procedure for trajectory optimization, with different modes using complementary methods of modeling collision costs and collision avoidance gradients. The primary stage queries the robot body's SDF, swept along the route to goal, at the obstacle point cloud, enabling swift local optimization of trajectories. The secondary stage infers the visible scene's SDF by aligning and composing the SDF representations of its constituents, providing better informed costs and gradients for trajectory optimization. The dual mode procedure combines the best of both stages, achieving a success rate of 98%, 14.4% higher than baseline with comparable amortized plan time on iGibson 2.0. We also demonstrate its effectiveness in adapting to real-world indoor scenarios.

Figures (4)

• Figure 1: Our proposed Dual Mode pipeline comprises the two pipelines shown above. Scene SDF pipeline (\ref{['fig:scene-sdf']}) infers the consolidated SDF of the robot's workspace by computing the mapping from the workspace to the domain of each individual SDF representation. On the other hand, the Robot SDF pipeline (\ref{['fig:robo-sdf']}) directly operates on the visible scene's point cloud by querying the robot body's SDF along the robot trajectory.
• Figure 2: Execution of the proposed pipelines is demonstrated for a given object placement at time step $t$ and after a displacement at time step $t+1$ (shown in the respective PC Alignment sections, where all three objects are displaced). The Robot SDF pipeline queries the robot body's SDF at the obstacle point cloud (black) to guide trajectory optimization. The Scene SDF pipeline assigns point cloud regions to each object instance, and then infers the mapping from the robot's workspace to the domain of each object's SDF representation. This mapping enables inferring the full SDF of the visible scene used to guide trajectory optimization. Red and blue spheres denote robot's start and goal positions, respectively.
• Figure 3: Operation of the Dual Mode pipeline is shown in a simulated indoor scenario (1), where the trajectory from Robot Body SDF pipeline is stuck in a local minima (2), triggering the Scene SDF pipeline which generates a collision-free trajectory (3). Subsequently, the Robot Body SDF pipeline picks-up from this trajectory and continues operation (4).
• Figure 4: We measure the impact of obstacle density in the robot's environment, using the scenes shown (left, obstacle densities noted on top-left corners). Results (right) show that the proposed Dual Mode pipeline exhibits the lowest deterioration in performance as the scene density increases. Plan Times (corresponding to successful simulations only) are plotted on $\log$-scale for convenience.