Module 01
Robot Brain / Embodied Intelligence / VLA
World models, action reasoning, VLA-style interfaces, and Real2Sim2Real learning that connect perception with physically grounded robot action.
Industry Application
Physical intelligence for real engineering systems.
PIRLab organizes industrial application around three modules: robot brains for embodied intelligence and VLA, industrial autonomous robots for construction automation, and autonomous mobile robots and autonomous driving.
Application Stack
From 3D sensing to deployable automation.
The CV highlights a deployment-oriented agenda: robot world models for engineering automation, infrastructure automation, construction robotics, 3D vision, SLAM, and robot manipulation.
The modules below organize that work by industrial capability: how robot brains reason and act, how autonomous robots operate in construction and infrastructure environments, and how robots perceive and localize in dynamic driving-scale scenes.
Three Modules
Industrial applications are grouped by deployable robotic capability.
Module 02
Industrial Autonomous Robots / Construction Automation
Robotic brick laying and autonomous crane operation for construction-site automation and industrial robotic deployment.
Module 03
Autonomous Mobile Robots / Autonomous Driving
Localization, mapping, tracking, segmentation, odometry, and scene-flow perception for robots and vehicles operating in dynamic environments.
Module 01
Robot brain, embodied intelligence, and VLA.
Continual VLA
Stellar VLA for evolving robot skills
Continual skill knowledge helps embodied agents acquire new manipulation tasks while retaining earlier capabilities.
Flow world model
RoboFlow4D for real-time action guidance
Predicted 4D flow gives the robot brain an explicit model of how objects should move under a task instruction.
Vision-Sound-Language-Action
HEAR for sound-centric manipulation
Streaming sound, causal memory, and multimodal reasoning let robots react to acoustic events during execution.
VLA
Continual skill knowledge
VLA systems need to accumulate task and skill knowledge across deployments without forgetting earlier behaviours.
World models
Flow as physical imagination
4D flow prediction lets robots anticipate scene evolution before committing to contact-rich actions.
Multimodal control
Seeing, hearing, and acting
Sound-aware policies extend embodied intelligence beyond static visual evidence and into temporal event understanding.
Module 02
Industrial autonomous robots and construction automation.
Brick laying robot
Autonomous robotic brick laying
The brick-laying robot demonstrates construction manipulation, site-aware motion, and task execution for physical automation workflows.
Autonomous crane
Autonomous crane operation
The crane demonstration focuses on autonomous lifting, perception-driven control, and industrial robot coordination in construction environments.
Deployment target
Construction-site autonomy
These systems move industrial robotics from structured lab tasks toward site-scale construction processes with heavy equipment and physical materials.
Robot capability
Manipulation and lifting
Brick laying and crane operation cover complementary capabilities: precise contact-rich manipulation and large-scale autonomous material handling.
Physical intelligence
Perception to action
3D perception, world modelling, and robot control are combined so construction robots can reason about geometry, tools, materials, and constraints.
Module 03
Autonomous mobile robots and autonomous driving.
2D-3D localization
Image-LiDAR registration for vehicle localization
Image-LiDAR registration, LiDAR odometry, segmentation, tracking, and scene flow create a robust perception layer for mobile platforms.
LiDAR odometry
Large-scale 3D localization and motion estimation
Deep LiDAR odometry and point-cloud registration support reliable state estimation for mobile robots and autonomous driving systems.
Perception stack
Localization, tracking, and scene flow
2D-3D registration, multi-object tracking, LiDAR odometry, and 3D scene flow form a practical perception stack for dynamic environments.
Road autonomy
Road-surface reconstruction
Multimodal road data, explicit meshes, and implicit encoding support large-scale road-surface mapping and maintenance-oriented analysis.
Robust deployment
Dynamic scene understanding
Segmentation, occupancy refinement, and motion prediction help robots reason about moving objects and changing road environments.