Object- and Task-Aware Active Perception for Safe Navigation of Mobile Manipulators in Semi-Static Environments
A mobile manipulator that points its hand camera at blind corners and changed objects along its planned path completed all nine test loops in cluttered rooms without a collision, while four baseline viewing strategies completed three to six, a preprint from the University of Toronto, Epson Canada, TUM and Simon Fraser University reports. In the hardest scene its map found 73.0% of obstacle voxels against about 31% for the baselines.
Can a mobile manipulator use its arm-mounted camera, while driving, to look where map errors would most endanger its current route, and does that make navigation in cluttered, slowly changing spaces safer?
Warehouses, factories and offices change between robot visits: boxes move, poles appear. Navigation stacks usually assume a static map, and a base camera with a narrow field of view misses obstacles behind blind corners until the robot is close, which leads to collisions, stops or needless replanning.
Fixed hand-camera poses, a camera that looks ahead along the path, random views, or next-best-view planners that maximise map information without regard to the route (NBV-Map). Object-aware semi-static mapping exists but is rarely coupled to viewpoint control.
A probabilistic next-best-view objective weights map uncertainty by how close a region is to the planned path and by whether objects there are likely to have moved (task-dependent map entropy). The view goal is folded into hierarchical-task model predictive control that ranks navigation above perception, so the arm's spare degrees of freedom aim the camera without hurting base tracking. Collision avoidance uses control-barrier constraints on spheres covering the robot. Hardware: a Universal Robots UR10 on a Clearpath Ridgeback base with two Orbbec Femto Bolt RGB-D cameras, all computed on an onboard laptop.
Across nine sequential loops (three scenes, three loops each) the method completed 100%; Lookahead 66.7%, Fixed, NBV-Map and Random 33.3% each. In the easy scene it had the best precision (87.4%), recall (70.9%) and false-positive rate (1.7%). In the L-shaped-wall scene its recall was 73.0% vs 28.4% to 31.2% for baselines, which all collided three times; in the office scene recall was 70.0% vs a best baseline of 58.3%. The extra coverage came with lower precision and higher false positives in the hard scenes (FPR 8.4% and 8.6%).
This is a preprint. The real-world benchmark is small (three indoor scenes, nine loops per method) with collisions triggering map resets. The authors attribute the lower precision in hard scenes to localisation drift and sensor noise. Scene-change behaviour is shown in controlled simulation and demonstration sequences rather than measured at scale, and the setup needs an arm with a camera, not standard on AMRs.
Relevant to makers of mobile manipulators for logistics and factory work (UR arms on AMR bases are a common integrator build) and to camera suppliers such as Orbbec. Epson's participation points to industrial interest. The method uses hardware that exists today, so warehouse and lab-automation deployments could adopt the idea within one to two years, mainly for safety in mixed, changing aisles.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.