Publications

A policy-optimization method that sharpens the visual-dynamics prediction inside video-based action models, improving how the policy anticipates how the scene will change before acting.

A practical, deployment-oriented Vision-Language-Action foundation model for robotics, designed to be trained and served at scale across pretraining, post-training, and reinforcement learning.

A benchmark of 100 fine-grained, detail-oriented manipulation tasks for stress-testing embodied AI agents on the kind of precise, long-tail behaviors that coarse benchmarks miss.

A single model that handles both multimodal understanding and robot control. Phased alignment training plus a mixture-of-experts design curb the "spurious forgetting" and task interference that usually appear when the two objectives are trained jointly.

Scales a transformer diffusion policy from 10M to 1B parameters. Affine feature-conditioning and an unmasking strategy tame the training instability that otherwise appears as the network grows, yielding large gains on simulated and real manipulation.

A family of compact VLA models that skip large-scale robot pretraining and attach a diffusion action decoder during fine-tuning, giving faster inference and better data efficiency for real-world deployment.

Equips VLA models with a "chain-of-affordance" reasoning step — explicitly reasoning about object affordances and reachable interactions before predicting actions, which improves manipulation success.

A coverage-path planner that combines an iterative potential field with segmented trajectory optimization and an incremental k-d tree, producing uniform, lightweight coverage for mobile robots.

Uses multi-view constraints from an RGB video stream — via scale-aware monocular SLAM and an object-level pose-graph optimizer — to recover robust, absolute category-level object pose and size without relying on accurate depth.

A sparse-convolution framework for semantic segmentation of LiDAR point clouds, the method underlying our S3Net / AF2-S3Net line of work on large-scale 3D scene understanding.

Extends scale-drift-aware pose graph relaxation to the case where the relative scale between consecutive monocular frames is unknown — as happens when monocular SLAM re-initializes with no reliable overlap — recovering globally consistent trajectories across scale jumps.

We present a novel semantic enhanced LiDAR-based localization algorithm that considers a robust combination of semantic and non-semantic information, enabling adaptive scene-dependent localization behaviour.

Inspired by spherical multi-view scanners, we propose a novel sampling model called Spotlights to represent a 3D shape as a compact 1D array of depth values. It simulates the configuration of cameras evenly distributed on a sphere, where each virtual camera casts light rays from its principal point through sample points on a small concentric spherical cap to probe for the possible intersections with the object surrounded by the sphere.

We optimized the network structure from S3Net and pushed our performance even further, and achieved state-of-the-art result on public dataset semanticKITTI dataset (sigle scan, named as AF2S3Net).

We proposed a real-time blazing fast Lite Harmonic Dense Block powered LiDAR point cloud segmentation network on spherical projected rangem map, and achieved state-of-the-art result on public dataset semanticKITTI dataset (sigle scan, named as LHDSegNet).

We proposed a novel sparse 3D convolutional neural network framework to tackle the large scale real-world point cloud semantic semantic segmentation chellenges and achieved state-of-the-art result on public dataset semanticKITTI dataset (named as kyber_HW).

We proposed a novel sparse 3D convolutional neural network framework to tackle the large scale real-world semantic scene completion chellenges and achieved state-of-the-art result on public dataset semanticKITTI dataset (named as Noah_Canada).

we propose a framework to inform and guide policy learning with augmented attention representations, demonstrating outstanding convergence speeds and stability for self-driving control.

We incorporate dense depth prediction and propose a novel deep learning module to improve the robustness and precision for the traditional Direct Sparse Odometry method.

Proposed a deep learning system enables safe and autonomous coral reef navigation in underwater environments. We validate our approach using an untethered and fully autonomous robot swimming through coral reef in the open ocean.

We modeled underwater visual navigation quality and turn it into training label for a deep synthetic network to guide robot swim along paths with low probability of visual tracking lost.

We proposed a pipeline to weight residual pattern based on the residual point image local context to simulate the feature encoding and matching process in in-direct approaches. We are aiming to reduce the non-convexity introduced by including original image in energy function.

We applied sparse convolution and transpose convolution on raw Kitti Velodyne point cloud data to predict dense semantic segmentation of BEV masks.

Learn to follow the lane in a complex environment with DQN and DDPG, augumented with the latent semantic input with raw rgb image. Experimental results demonstrated outstanding convergence rate and stability.

We proposed a 3D topological graph map to fit for large scale 3D Lidar localization.

We show the utility of a hybrid segmentation module to extract key LiDAR points for landmark-based localization, built upon a novel topological map structure.

We propose an end-to-end convolutional network architecture for accurately segmenting road masks at extended ranges from LiDAR point clouds in real-time.

We exploit depth prediction as a candidate prior for the coarse initialization, tracking, and marginalization steps of the direct visual odometry system, enabling the second-order op-timizer to converge faster to a precise global minimum. In addition, the given depth prior supports large baseline stereo scenarios, maintaining robust pose estimation against chal-lenging motions such as in-place rotation. We further refine our pose estimation with semi-online loop closure. The experiments on KITTI demonstrate that our proposed method achieves state-of-the-art performance compared to both traditional direct visual odometry and learning-based counterparts