Autonomous Door Opening — CALL-M Mobile Manipulator
Overview
At the CNRS-AIST Joint Robotics Laboratory in Tsukuba, I worked across the complete autonomous door-opening pipeline for CALL-M, a mobile manipulator built for indoor service robotics. After the robot pose is initialized in RViz, CALL-M searches for a door, approaches it autonomously, detects and localizes the handle, aligns the mobile base, positions the manipulator, grasps and actuates the handle, and attempts to open the door without further operator intervention. The system was evaluated in 20 real-robot trials and documented in a submitted research paper.
Seven-Stage Pipeline
- Door Search
- Approach
- Handle Detection
- 3D Estimation
- Fine Alignment
- Pre-Grasp
- Contact, Grasp, Handle Actuation, and Door Opening
A YOLO-based door detector (trained on 4,656 images) triggers the autonomous approach; Nav2 brings the robot to roughly 0.80 m from the door; RGB-D and point-cloud data localize the handle in 3D and TF carries that target across the camera, base, manipulator, and map frames; small corrective motions of the omnidirectional TriOrb base align the handle within the manipulator’s workspace before the arm executes a pre-grasp, grasp, and door-opening Cartesian sequence.
My Work — End-to-End System Development
I worked across the complete project rather than a single subsystem: ROS 2 system integration, door-search behavior, the door- and handle-detection pipeline, RGB-D/point-cloud handle localization, TF transforms, Nav2 integration and base alignment, manipulator positioning and pre-grasp planning, gripper control and handle actuation, real-robot experimentation, stage-level evaluation, failure analysis, and contributing to the resulting paper.
Results
The system was evaluated in 20 real-robot trials across three scenarios.
Nominal (10 trials, 0.35 m initial distance): Door Search, Approach, Handle Detection, 3D Estimation, Fine Alignment, and Pre-Grasp each succeeded 10/10; complete door opening succeeded 8/10.
Initial distance variation (5 trials, 2.15–10.10 m): 4/5 end-to-end successes — the 10.10 m trial failed during approach after localization drift reached the edge of the global costmap.
Initial end-effector height variation (5 trials, 0.34–0.68 m): 1/5 end-to-end successes, succeeding only at the nominal 0.44 m height; other heights produced alignment, grasp, or actuation failures.
Key Finding
Perception, localization, and navigation were highly repeatable across all three scenarios. Nearly all end-to-end failures occurred during physical interaction with the handle — residual alignment error, grasp depth, force-limit activation, and insufficient latch displacement. Closed-loop pose correction and force-aware manipulation are the clearest directions for improvement.