UmU_Spatial_Lab_Robotics_Traversability
Abstract
Classical robot navigation uses dead-reckoning techniques like odometry to estimate
position, heading, and distance travelled relative to a static map. While it is widely used, this method is also outdated, odometry drifts over time, and even minor physical or computational errors can add up to a significant misalignment over a long trajectory. Deep learning offers a different approach that depends less on these static assumptions.
Our research looks at end-to-end neural models that take vision (sequence of RGB frames i.e. video, not a single image) to pick out stable visual landmarks for localisation and to
predict a feasible trajectory, without relying on explicit odometry or a pre-built static map.
The main gap noticed across the papers in the literature is high-accuracy navigation in indoor/closed environments with reliable obstacle avoidance for both static obstacles
(walls, furniture) and dynamic obstacles (people, moving objects), one of the four papers reviewed explicitly test dynamic-obstacle avoidance. SwarmDiffusion [4] largely solves this issue, but it isn’t a single model that takes in an RGB image and directly outputs waypoints along with their distances in world coordinates. Parts of the problem have been solved individually across different papers but a single, unified implementation that performs well with low latency has not yet been achieved.
Relevant papers are GNM Shah et al. [1] and SwarmDiffusion Zhura et al. [4], as they address the same indoor navigation problem of high-accuracy maneuvers. However, GNM and SwarmDiffusion do not use the more advanced diffusion algorithms to reach goals, which are better implemented in ViNT Shah et al. [2] and NoMaD Sridhar et al. [3].
So, our research questions are:
RQ1: Can a traversability map be given to a diffusion model to get waypoints and their distances in world coordinates?
RQ2: How can LLMs be used to understand commands and turn them into navigation goals for these models?
To investigate this we will use the dataset mentioned in GNM [1], ViNT [2], and NoMaD [3]. And the SwarmDiffusion [4] dataset collected from a mounting a camera on a Unitree Go1 and a UAV, gathering a total of 8000 samples and pairing them with their respective annotated trajectories using a self-supervised algorithm.
