Robotics Briefing /
Robots that remember, recover, and feel.
A useful robot needs to remember what happened, stay upright when conditions change, and feel objects slipping through its fingers. Today’s papers test those abilities on physical machines. An industrial robot launch and a new dexterity benchmark show how the supporting hardware and evaluation tools are developing.
Your quick takeaways
- MessyMem lets a mobile robot reuse discoveries across tasks, including which drawer is locked and where a matching object was seen.
- ResSafe reduced humanoid falls across four physical test conditions. Touch2Trace used fingertip pressure to guide an unfamiliar cable through a robot hand.
- Universal Robots introduced Gen 7 at IMTS, with easier sensor connections and force control. Bench2Dex adds a shared simulation benchmark for two-handed manipulation.
Papers
MessyMem: remembering what a robot learns by doing
Anuva Banwasi and colleagues, Stanford • Submitted September 14 • Authors report CoRL 2026 acceptance; simulation and physical tests
A robot searching an office should remember that a drawer is locked and where it already found a particular cup. MessyMem stores object locations, records what interactions reveal, and links those records to images. Later tasks can retrieve the relevant experience, including visual details that a simple object label would miss.
In a simulated sequence of 25 tasks spanning more than three hours, it reached 80% task progress, compared with 65.2% for the strongest version missing one memory component. Progress gives credit for partial completion; it is not an 80% rate of finishing entire tasks.
On a physical mobile manipulator, the complete system achieved full progress in all five office-search trials and all six sock-pairing trials. The comparisons shared planning and movement tools, helping isolate the contribution of memory.
The broader promise is fewer repeated searches and failed attempts during extended operation. Physical evidence remains small, however: the robot used wall markers for localization and a predefined set of object categories. Tracking changes caused by people remains future work.
ResSafe: a second controller helps humanoids keep their balance
Gechen Qu and colleagues, UC Berkeley and UCLA • Submitted September 14 • Preprint; simulation and physical tests
A humanoid can reproduce a movement accurately and still fall when extra weight changes its balance. ResSafe first trains a movement controller, then holds it fixed while a second controller learns small corrective actions. Training exposes the robot to disturbances and varying payloads.
Across four physical Unitree G1 test conditions, with ten trials each, the reported results sum to three falls in 40 attempts. A stronger comparison controller trained with varying payloads fell nine times in 40 attempts. ResSafe did not win every condition: in one, that comparison had no falls while ResSafe had two.
Separating movement from balance correction could make it easier to improve robustness without retraining every skill around competing objectives. Tests with multiple versions of the underlying controller also examined whether the corrections remain useful as that controller changes.
The filter provides no hard mathematical safety guarantee, and falls still occurred. It relies on body and joint state without external visual awareness. These results concern balance during tested motions; broader workplace safety requires evidence beyond this evaluation.
Touch2Trace: guiding a cable through a hand by touch
Matteo Grimaldi and colleagues, Analog Devices • Submitted September 14 • Authors report CoRL 2026 acceptance; physical hand experiments
Flexible cables shift as they move, making a steady grip difficult. Touch2Trace uses fingertip pressure maps and joint positions to control repeated pinching and curling movements on a Tesollo hand. Its movement model learned from 12 demonstrations totaling 10.1 minutes, using a touch encoder pretrained on separate contact data.
On an unfamiliar straight Ethernet cable, mean tracing distance reached 20.1 centimeters, compared with 0.2 centimeters using joint positions alone. Across 30 scored trials, 93% reached at least ten centimeters and 57% reached twenty. Aborted starts lasting under five seconds were excluded.
The results show how fast, detailed contact feedback can sustain manipulation of a deformable object. The controller ran at 60 updates per second, using a short history of pressure readings to adjust its movements.
Transfer was uneven: only 47% of trials reached ten centimeters on a thinner USB cable. This was a fixed-hand tracing task, with a limited set of cables. Complete wiring jobs would also require locating cables, moving the arm, and handling connectors.
News
Universal Robots launches Gen 7 at IMTS
Launched September 14 in Chicago • Regional announcement September 15 • Company product announcement
Universal Robots introduced three g-Series arms alongside a redesigned controller, teaching pendant, and tool connection. The connection carries data, power, and safety signals to cameras and other sensors. Built-in force sensing and impedance control let the arm regulate how firmly it interacts with objects.
The company reports 40% more compute in a controller with a 30% smaller footprint than previous generations. Its PolyScope X software provides open interfaces and ROS 2 communication for connecting external AI processing. The controller also supports existing e-Series and UR Series arms.
These changes could reduce the wiring and integration work needed for vision-guided handling and force-sensitive assembly. The announcement does not provide independently measured productivity gains or sustained deployment results for the new platform. Those outcomes will establish how much the easier connections and added compute improve factory work.
Highlights
Bench2Dex: a common test for two-handed robot skills
Zhenjie Yang and colleagues, Shanghai Jiao Tong, Fudan and collaborators • Submitted September 14 • Simulation benchmark; public code
Comparing robot hands is difficult when each system uses different tasks, sensors, and success rules. Bench2Dex brings together 26 two-handed manipulation tasks, 12 hand designs, and approximately 1,300 human-controlled demonstrations collected in simulation. It includes tests that change conditions and checks whether completed task states remain stable.
The public repository contains collection, replay, training, and evaluation code, with links to data and model checkpoints. That gives researchers a shared starting point for comparing how vision and touch contribute to dexterity.
Its touch representation comes from simulated contact geometry and does not reproduce a particular physical sensor. Benchmark scores therefore cannot establish hardware performance or reliable transfer to real touch sensors. The contribution is a more consistent way to compare systems before physical testing.
Read this first
Read MessyMem. Remembering a locked drawer or a sock pattern makes the value of persistent robot memory easy to understand. Its controlled comparisons also show which kinds of memory help, while the small physical evaluation keeps the evidence in perspective.
Read the paper