Robotics Briefing /
Gentler contact, faster collision models, and routes that respect the worksite.
Reliable robots need more than a good destination or a plausible-looking scene. They need to respond when contact changes their motion, model the shapes that can block a path, and know when an otherwise clear route is off limits. Today’s research tackles those practical constraints, alongside a logistics dataset that preserves operational failures and a funded study of assistance for older adults.
Your quick takeaways
- A learned contact reflex improved tight peg insertion on hardware, but even its best result in the hardest setting was 58% success.
- I2CD prepared collision geometry for twenty objects in eleven seconds. Physical pick-and-place success was slightly lower than the slower comparison.
- CORNAV combines blueprints and work schedules with robot perception. Its main quantitative results come from recorded-data planning tests, not a large live deployment.
Papers
A learned reflex lets contact guide the robot
Quan Nguyen, Yunho Kim and Joonho Lee, Neuromeka • October 2 submission; newly listed October 5 • Preprint
A peg meeting the rim of a hole needs to yield and search, rather than push harder toward an imperfect position estimate. This work trains a low-level reflex in simulation using three simple interactions: a spring, a plane and a rail. It then freezes that reflex beneath the controller that chooses the task’s next movement.
The reflex uses the robot’s recent internal state, without direct force or geometry measurements. It adjusts joint targets so contact can constrain motion while preserving the intended action.
On physical peg insertion with 0.02-millimeter tolerance and a two-millimeter location error, the reflex-based system succeeded in twenty-eight of fifty trials, versus nine of fifty for the position-control comparison. At zero imposed offset, the counts were twenty-nine and eleven. These are meaningful gains, but substantial failures remain.
Mean estimated contact force was also lower. Force values came from joint-torque residuals with brief spikes filtered, rather than a direct force-sensor measurement.
Each execution layer had its own separately trained insertion policy, so this compares complete systems rather than swapping controllers under identical commands. The reflex also needs an explicit signal telling it when to yield, and its training primitives do not cover coupled rotation and translation, such as opening a door with a rigid grasp.
I2CD: turning an image into geometry a robot can plan around
Qian Wang and colleagues, Yale University • October 2 submission; newly listed October 5 • Preprint with physical tests
A visually convincing 3D model is not necessarily suitable for collision checking. Simulators often need objects represented as collections of simple solid shapes. Converting detailed generated meshes into that form can be slow, and a simulator may silently substitute a shape that behaves differently.
I2CD predicts the collision representation directly from an image. It keeps a pretrained image-to-3D model fixed and trains a smaller component to produce a set of convex parts—shapes without inward dents that collision algorithms can handle efficiently.
In a tabletop scene containing twenty objects, I2CD produced planner-ready geometry in eleven seconds, compared with 328 seconds for SAM3D followed by CoACD decomposition. A physical xArm7 then completed eighty-five of one hundred pick-and-place trials with I2CD, versus ninety with that comparison.
The test held segmentation, placement of objects, planning and hardware fixed while changing geometry. Success required completing the planned sequence without collision, so the speed gain did not come with a demonstrated improvement in task success.
The result could shorten the wait between observing a scene and planning movement. However, it covers one tabletop scene, and the predicted geometry can miss thin features or smooth away important detail. Tasks involving repeated sliding or pushing may be more sensitive to those errors than the evaluated pick-and-place task.
CORNAV: a clear corridor can still be the wrong route
Parastoo Ali Pour and colleagues, UC Irvine and IIT Bombay • October 2 submission; newly listed October 5 • Preprint
On a construction site, a route can be physically open but restricted because work is underway. CORNAV combines robot observations with architectural drawings and project schedules. Blueprints help identify rooms and permanent features; schedules turn work areas into mandatory exclusions or higher-cost regions that the planner prefers to avoid.
The system also uses a language-model check to identify hazardous requests when a schedule labels an area too permissively. The path planner then applies the resulting constraints.
Across fifty-four requests evaluated on recorded sensor data, the full system’s reported task-success score was 72.2%, versus 13% without blueprint grounding. Its planned paths had no mandatory-zone violations in the evaluation. Those percentages describe offline evaluation using surveys from an office and construction site.
The authors separately report twelve physical experiments using Go2 and G1 robots, with demonstrations of reaching destinations or declining infeasible routes. That small live study should not be confused with the larger recorded-data comparison.
The broader contribution is making site knowledge part of navigation. Reliability still depends on aligning drawings with observations, maintaining accurate schedules and recognizing hazards. Zero violations of the represented zones does not establish general safety around a changing worksite.
News
Hello Robot receives funding to test assistance in homes and care settings
September 29 announcement • Catch-up • Research grant
Hello Robot announced nearly $3 million over three years from the National Institute on Aging for work with the University of Illinois Urbana-Champaign. The project will evaluate Stretch 4 assistance for older adults with mild cognitive impairment or early Alzheimer’s dementia, including reminders, finding objects, social engagement and daily routines.
The planned research includes sixty older adults, care partners and professional staff, with extended evaluations in homes and care settings. The project began September 22; coverage published this weekend concerns the earlier award.
The useful development is support for studying reliability, acceptance and practical benefit over time. This is a research program, not evidence that Stretch already improves health outcomes or replaces care staff.
Highlights
MoRoOp: robot logistics data that keeps the delays and failures
Jan-Felix Klein and Yongkuk Jeong, KTH • October 1 submission; newly listed October 5 • Dataset paper
MoRoOp links transport requests to dispatch decisions and robot telemetry across nine eight-hour laboratory shifts. It contains 1,382 jobs, 4,815 operations and 19,352 dispatch events. Valid records of obstructed navigation, delays and unsuccessful operations remain included.
That connection helps explain whether a late job arose from dispatching, movement or another operational step. Actual data files are available, with both raw and cleaned robot-state tables.
The setting used one robot and generated jobs in a laboratory. Station processing times were assigned waits, not measurements of people or machines performing production work. It offers a useful operational trace, without representing the full variability of an industrial fleet.
Read this first
Read I2CD’s physical evaluation. Its eleven-second geometry preparation and slightly lower task success make the practical tradeoff between speed and accuracy easy to see.
Read the paper