Humans navigate large, open-world environments not by maintaining a globally consistent metric map, but by forming cognitive maps—topological representations that support robust relocalization even under severe perceptual changes. This paradigm is inherently resilient to tracking failures, dynamic environments, and long-term appearance variations. A fundamental enabler of this robustness is sequential hypothesis testing (SHT), which humans implicitly perform to disambiguate perceptually aliased places over time.

Despite its importance, most robotic systems either (i) avoid aliasing through conservative data-association thresholds, or (ii) apply filtering only in a discrete place space constructed from keyframes. Such discretization restricts the hypothesis space, prevents reasoning in between stored views, and leads to brittle relocalization when observations do not perfectly overlap past frames.

We propose a pose-aware online topological mapping system that performs SHT directly in continuous SE(3) space. Our approach maintains and updates a multi-hypothesis belief over the robot pose and the map topology, allowing robust relocalization and loop closure without relying on global metric consistency. Addressing continuous-domain SHT introduces two major challenges: (1) how to represent and update distributions in SE(3) efficiently, and (2) how to manage map uncertainty—as each perceptual aliasing event can spawn multiple plausible map branches, whose number can grow rapidly if not controlled.

We introduce a Gaussian-mixture–based representation and a principled branch management strategy that together enable real-time inference, scalable map maintenance, and robust operation in dynamic, ambiguous, or previously unseen environments. This work provides the first practical continuous-space SHT framework for topological mapping, bringing robots closer to the reliability and flexibility of human navigation.

We propose a data-driven reinforcement learning framework for social navigation that turns real-world human–robot navigation logs into replayable "imagination environments." This approach allows robots to learn navigation policies directly from logged RGB-D data and human trajectories, eliminating the need for conventional simulators.

The system reconstructs bird's-eye-view (BEV) maps from RGB-D observations and replays recorded human trajectories within these environments. The robot can then train and refine its navigation policies by interacting with these reconstructed scenarios, learning socially-aware behaviors from real-world data.

This framework bridges the gap between simulation-based training and real-world deployment, enabling robots to learn from actual human–robot interactions without the complexities and limitations of traditional simulation environments.

This project leverages vision–language models to enable zero-shot visual-language navigation without the need for explicit policy training. The system translates natural-language navigation instructions into differentiable cost functions defined over bird's-eye-view (BEV) scene representations.

By optimizing or selecting trajectories that minimize these learned cost functions, the robot can follow complex natural language instructions while navigating through real-world environments. This approach eliminates the traditional requirement for large-scale policy training on navigation datasets.

The method enables flexible, instruction-following navigation that can adapt to diverse language commands and environmental contexts, bridging vision, language, and action in a unified framework.

This research focuses on using offline reinforcement learning to fine-tune vision–language–action (VLA) models using real robot interaction datasets collected in laboratory environments. Rather than training from scratch, we leverage pre-trained VLA models and adapt them to specific robotic tasks through offline RL.

By learning from logged interaction data, the system can improve task performance without requiring extensive online exploration, which can be costly and time-consuming in real-world robotics. This approach enables safer and more efficient learning from limited real-world data.

The method demonstrates how large-scale pre-trained models can be effectively specialized for robotic manipulation and navigation tasks through principled offline learning, accelerating the deployment of capable robot systems in real-world environments.