This project develops LatentAM, an online 3D Gaussian Splatting mapping framework that builds scalable latent feature maps for open-vocabulary robotic perception. We propose an online dictionary learning approach that is both model-agnostic and pretraining-free, enabling plug-and-play integration with different Vision-Language Models at test time. To support large-scale operation, we introduce a local-global map management strategy that stores the global map on the CPU while optimizing only an active local map on the GPU. Experiments on public benchmarks and a custom large-scale dataset demonstrate that LatentAM achieves high-fidelity feature reconstruction, near-real-time performance, and scalable open-vocabulary 3D mapping.

This project develops CORD and DeepCORD, a family of distributed geometric optimization algorithms for scalable multi-robot perception backends. CORD (RSS'26) formulates distributed pose graph optimization as a damped second-order Riemannian dynamical system on Lie groups, where pose variables evolve with inertia and damping toward first-order critical points of the original optimization problem. This formulation enables fully parallel distributed updates with minimal communication and uses velocity-based neighbor prediction to improve robustness under delayed or asynchronous communication. Building on CORD, DeepCORD unfolds the continuous Riemannian dynamics into differentiable iterations and learns an adaptive local feedback policy that dynamically selects solver parameters from local optimization and communication context. Experiments on large-scale SE(3) pose graph optimization and SL(4) projective submap alignment demonstrate improved convergence and robustness over hand-tuned distributed baselines across synchronous and asynchronous regimes.

This project develops MISO (RSS'25), a hierarchical optimization approach that leverages multiresolution submaps to achieve efficient and scalable neural implicit reconstruction. For local SLAM within each submap, we develop a hierarchical optimization scheme with learned initialization that substantially reduces the time needed to optimize the implicit submap features. To correct estimation drift globally, we develop a hierarchical method to align and fuse the multiresolution submaps, leading to substantial acceleration by avoiding the need to decode the full scene geometry. MISO significantly improves computational efficiency and estimation accuracy of neural signed distance function (SDF) SLAM on large-scale real-world benchmarks.

This project develops the first approach for distributed multi-robot dense metric-semantic mapping. In (Tian et al., T-RO'22), we present Kimera-Multi, a multi-robot system that (i) is robust to spurious loop closures resulting from incorrect place recognition, (ii) is fully distributed and only relies on local communication, and (iii) builds a globally consistent metric-semantic 3D mesh model of the environment in real-time, where faces of the mesh are annotated with semantic labels. Kimera-Multi has been demonstrated in real-world field experiments with up to 8 robots traversing up to 8km (Tian et al., IROS'23). 

IEEE Transactions on Robotics King-Sun Fu Memorial Best Paper Award, 2022.
Honorable Mention for MIT Open Data Prize, 2023.

This project develops a suite of distributed geometric optimization algorithms, the backbone of modern collaborative simultaneous localization and mapping (CSLAM) and camera network localization (CNL) systems. In (Tian et al., T-RO'21), we developed first certifiably correct algorithm for distributed pose graph optimization (PGO). In (Tian et al. RA-L'20), we extend the approach to handle delayed communication while maintaining provable convergence. In (Tian et al. T-RO'23), we further leverage spectral graph theoretic tools to achieve a significant speed up in the algorithm convergence.

Honorable Mention for IEEE Robotics and Automation Letter Best Paper Award, 2020.
Honorable Mention for IEEE Transactions on Robotics King-Sun Fu Memorial Best Paper Award, 2021.