Robotics Research Intern

Overview

At the CNRS-AIST Joint Robotics Laboratory in Tsukuba, I worked on autonomous mobile manipulation using CALL-M, a research mobile manipulator built for operation in human-centered indoor environments.

My work centered on two research projects:

• Autonomous mapping and exploration in unknown, dynamic environments.
• Autonomous door opening through integrated perception, navigation, and manipulation.

Across both projects, I worked on ROS 2 architecture, LiDAR/RGB-D perception, point clouds, TF, SLAM, frontier exploration, Nav2, dynamic-obstacle filtering, mobile-base control, manipulation, distributed execution, real-robot testing, failure analysis, and research writing.

CALL-M

CALL-M combines:

• TriOrb omnidirectional mobile base.
• UR5e manipulator.
• Robotiq 2F-85 gripper.
• RPLIDAR sensors.
• ZED Mini stereo cameras.
• Intel RealSense D435i.
• Onboard Intel and NVIDIA compute.

Project 1 — Autonomous Mapping

I developed a modular ROS 2 architecture that lets CALL-M autonomously explore an unknown environment, build an occupancy map, supervise navigation goals, recover from failures, filter moving objects, and stop when exploration is complete.

My main contributions included:

• Multimodal range-sensor integration.
• Frontier extraction and target evaluation.
• Nav2 goal dispatch and supervision.
• Timeout and failure recovery.
• Spatial memory of failed targets.
• Mission-progress tracking.
• Mission-completion logic.
• Static–dynamic obstacle separation.
• Object clustering, tracking, and velocity estimation.
• Simulation and physical-robot validation.

The exploration architecture completed 10/10 simulated missions compared with 7/10 for a nearest-frontier baseline, and all 5/5 physical mapping sessions reached autonomous completion.

Project 2 — Autonomous Door Opening

I also worked on a mobile-manipulation pipeline for autonomous door opening. The system:

• Detects a door using machine-learning-based perception.
• Approaches it using Nav2.
• Localizes the handle in 3D from RGB-D data.
• Transforms the target through ROS 2 TF.
• Aligns the omnidirectional base.
• Moves the UR5e to a pre-grasp pose.
• Grasps and actuates the handle.
• Executes the door-opening motion.

The pipeline was tested in 20 real-robot trials. In the 10 nominal trials, perception, navigation, 3D localization, alignment, and pre-grasp each succeeded 10/10; complete door opening succeeded 8/10. The experiments identified physical contact and manipulation geometry as the main remaining limitations.

Research Output

• Authored 2 research papers as first author and contributed to a third as co-author during the internship.

My AIST work resulted in three research-paper efforts:

1. A Supervised ROS 2 Architecture for Autonomous Mapping: Modular Exploration-Management Framework — First author — Submitted to IEEE/SICE SII 2027
2. A ROS 2 Static–Dynamic Obstacle Filtering Framework for Robust Mapping in Dynamic Environments — In Preparation
3. An Integrated Perception, Navigation, and Manipulation Pipeline for Semi-Autonomous Door Opening with a Mobile Manipulator — Second author — Submitted

Related Project & Research Pages

• View Project — Autonomous Mapping (CALL-M)
• View Research — A Supervised ROS 2 Architecture for Autonomous Mapping
• View Project — Autonomous Door Opening (CALL-M)
• View Research — Semi-Autonomous Door Opening Pipeline

Closing

AIST gave me experience building complete robotic systems rather than isolated algorithms. I worked from sensing and coordinate transforms through autonomous decision-making, navigation, manipulation, distributed software, and physical validation.