Robotics Briefing /
Precise assembly without cameras. Robot memory put to work.
A robot can find a tight opening through changes in its own motion. Remembering where objects belong can help it restore a room, but picking them up remains difficult. Today’s research also tackles a less visible problem: how long a moving robot should spend planning before it needs fresh feedback.
Your quick takeaways
- MIT’s blind assembly policy completed 173 of 237 physical peg insertions after training entirely in simulation. It starts with the part already touching the fixture.
- FOUND-IT++ connects a robot’s visual memory to navigation and grasping. It located 18 of 20 moved objects, but picked up only 9 of those 18.
- Learning when to replan improved simulated navigation and drone control. On a simpler manipulation task, a fixed planning budget worked just as well.
Papers
Blind Precise Assembly: joint motion can reveal whether a part fits
Antonia Bronars, Younghyo Park, Kushal Kedia and Pulkit Agrawal, MIT • October 16 preprint; October 19 arXiv listing
An arm’s joint sensors measure how it moves. When a held part meets resistance, that motion changes, providing clues about contact without a camera or dedicated touch sensor. This work asks whether those clues are enough for precise assembly.
The learned controller reads a short history of joint positions, velocities and previous actions. Training varies part shapes, grasps and contact conditions across simulated assemblies. It learns to search and correct alignment rather than following a fixed sweep.
On a physical Flexiv arm, one unchanged policy seated eight different pegs in 173 of 237 trials, or 73%. Three pegs were outside the training distribution, and clearances ranged from 0.5 to 1.5 millimeters per side. The policy received no camera image, force measurement or hole location. Gear meshing and nut threading were demonstrated in simulation only.
The result shows how much useful contact information existing joint sensors contain. It is a final insertion skill: every trial begins with the held part touching the fixture near the hole. Another system must get it there, and the controller cannot recover a dropped part far away.
FOUND-IT++: remembering a room is only part of putting it back
Dominic Maggio and colleagues, MIT and Samsung Research America • October 16 preprint; October 19 arXiv listing
A useful robot needs more than an inventory of objects. It needs to know where they were, what changed and how to reach them. FOUND-IT++ turns a single camera’s observations into a layered 3D map linking objects, traversable places and larger regions.
The system adds navigation clearance checks, saved mapping sessions and change detection to an earlier mapping method. On a mobile manipulator, that memory supports a complete sequence: inspect familiar locations, find displaced objects, navigate to them and attempt to return them.
Across ten rearrangement missions in five scenes, it detected all twenty missing objects and found eighteen at their new locations. Grasping succeeded for nine of those eighteen. Initial mapping was teleoperated; the rearrangement phase ran autonomously, with operator intervention after failures that would otherwise stop a mission.
This is useful evidence about connecting robot memory to action, rather than just producing attractive maps. It also exposes the remaining gap. Final approaches sometimes brought the robot too close to tables, and missing rapid depth feedback made grasping difficult. These component results should not be mistaken for a fully autonomous household success rate.
Robot metareasoning: when thinking longer becomes a liability
Lorenzo Magnino and colleagues, University of Cambridge and George Mason University • October 16 preprint; October 19 arXiv listing
Planning takes time while a robot and nearby obstacles keep moving. Longer computation can produce a better route, but it also means acting longer before incorporating new observations. A fixed planning interval cannot suit every situation.
This method learns to choose both how long to execute the current plan and how much computation to give the next one. It considers the robot’s state and statistics from the planner, shortening commitments near moving obstacles or difficult passages and allowing more time when deeper planning helps.
In a simulated drone task involving a suspended load, the adaptive controller succeeded in 80% of twenty trials while all tested fixed commitments failed. Physical RoboMaster navigation demonstrations also showed the learned timing behavior transferring from simulation. The drone result is not a physical flight success rate.
Adaptation was not universally better: on the simulated Push-T manipulation task, the method settled on a constant budget and matched the best fixed baseline. The experiments also used an A100 GPU. The contribution is evidence that controlling when to react can matter alongside improving the plan itself, with benefits depending on the task and computing setup.
News
ABB adds retraining tools to apparel sorting
October 15 announcement • Catch-up • Company claims
ABB has integrated AI Robot Trainer into its Fashion Inductor module, which feeds apparel items into sorting systems. The update combines pretrained fashion vision models, tools for retraining on a particular cell and a digital replica of the module’s software for setup and validation.
The practical issue is changing inventory: new products and packaging can upset a system that worked well at installation. ABB describes a workflow that monitors performance, detects anomalies, generates training data and deploys updated models.
ABB claims up to 75% less engineering and commissioning effort. Its stated throughput is up to 1,300 items per hour per cell, with picking accuracy targets above 99.5%. Those are company figures and targets, not independent deployment measurements. The announcement does not provide the comparison behind the engineering savings. Expansion to other vision applications through the first quarter of 2027 is a plan.
Highlights
mjfit makes small assembly clearances explicit in simulation
Public resource accompanying Blind Precise Assembly
A simulated hole can look correct while its collision geometry blocks a part that should fit. The public mjfit repository generates pegs, holes, threads and gears from geometric pieces that preserve their intended clearances. It includes source code, examples, tests and benchmarks under an MIT license. The accompanying project page offers interactive assembly simulations. This makes the underlying modeling problem inspectable, though accurate geometry alone does not guarantee accurate friction or real-world transfer.
Read this first
Explore the Blind Precise Assembly project page. Its physical trials and interactive simulations make the central idea easy to see: an arm can infer contact from its own motion. Keep the starting condition in view; the robot is already holding the part against the fixture.
Read the paper