Robotics Briefing /
Better robot decisions, from the next move to the full job.
A useful robot has to choose a good next move, apply enough force, and leave room to finish the job. Today’s papers test those decisions on real hardware and in simulation. A new warehouse announcement adds a named customer site, while leaving the robot’s sustained productivity an open question.
Your quick takeaways
- SeeQ helped a robot hang a shirt in 22 of 24 trials, up from 10 of 24 with the same base policy. It ranks possible actions using the current subtask.
- ForceTwin measures how doors and drawers resist movement. Its gains are strongest where springs and other mechanisms cause simpler controllers to stall.
- X Square says its dual-arm robot now handles parcels at lululemon’s Wuhan distribution center. The announcement does not disclose throughput or intervention rates.
Papers
SeeQ: choosing better actions during a longer task
Saksham Singh and colleagues, Carnegie Mellon University • Submitted September 18; listed September 21 • Preprint; physical robot tests
SeeQ adds an action evaluator to an existing robot policy. It identifies the current subtask, such as inserting a hanger into a sleeve, then scores eight possible sequences of actions. The robot executes the most promising option.
Across four physical tasks, each with 24 trials, the method improved every result. Shirt hanging rose from 10 to 22 successes, lid sealing from 10 to 15, grocery packing from 9 to 17, and Lego disassembly from 5 to 10. The comparisons kept the base policy’s weights fixed.
This shows that better action selection can improve a trained robot without changing the policy that proposes its movements. The evaluator still needs its own training: SeeQ uses broad robot data, subtask labels, and further training on each target task.
Its judgment can fail if it identifies the wrong subtask. It can also favor immediate progress that makes later steps harder, such as knocking something over while packing. Public code and a pretrained checkpoint are available; the results do not establish performance on tasks without further training.
ForceTwin: measuring the resistance a camera cannot see
Tim Engelbracht and colleagues, ETH Zurich and collaborators • Submitted September 18; listed September 21 • Preprint; physical robot tests
Two similar doors can require very different effort to open. ForceTwin records a person moving an object with a force-sensing gripper, then builds a digital model of its motion, friction, inertia, and internal mechanisms. Those measurements help a robot anticipate the force it needs.
Across nine object–robot pairings using Spot and a Franka FR3, the average fraction of the target opening reached within six seconds was 87.3%. Visual-and-language estimates reached 59.7%, and a model without force compensation reached 56.9%. These numbers measure progress toward an opening target, not the percentage of successful trials.
The largest gains came from strongly resisting objects. On the Franka metal-door test, ForceTwin reached 81.2% of the target, while both comparison methods remained below 2%. Each trial started with the handle already grasped.
The approach makes physical measurements useful for both control and simulation. It requires probing each object, and its model can become outdated when a drawer’s contents or a door closer changes. Spot could not operate the metal door within its actuation limits.
The project lists code and videos as forthcoming.
Dense packing: the gripper needs space, too
Tianhao Qin and colleagues, Worcester Polytechnic Institute • Submitted September 18; listed September 21 • Authors report ISRR 2026 acceptance; physical tests
Fitting an object into a box is only useful if the robot can place it there. This system plans with the object and gripper together, monitors force during lowering, then gives the released object a small push to close leftover gaps. Cameras rescan the container between placements.
On one seven-object ordering, the full system completed all 35 placement attempts across five physical trials. Removing both gripper geometry and the associated orientation constraint reduced that to seven of 35; fingers hit walls, and the planner proposed unreachable poses.
The paper also reports 95.6% space utilization for its optimized ordering. That metric measures the share of the object set’s volume packed fully inside. It does not measure the fraction of the box filled; another ordering scored 75.2%, showing how much sequence matters.
The result connects geometric planning with the constraints of an actual gripper. It assumes a known object library and one predefined grasp per object. That leaves packing unfamiliar products, choosing grasps, and handling a wider range of materials for future work.
News
X Square names a warehouse deployment at lululemon
X Square Robot • Announcement September 20; facility opened September 16 • Company-reported deployment
X Square says its QUANTA X1 Pro wheeled, dual-arm robot handles and organizes parcels in daily operations at lululemon’s Wuhan distribution center. The update gives the company a named operating site following its earlier logistics demonstrations.
Logistics partner SF Express separately confirms the center’s September 16 opening and describes unloading robots, mobile transport, sorting equipment, and RFID tracking. Its announcement does not name QUANTA X1 Pro. It corroborates the facility and surrounding automation, leaving the specific robot’s performance unverified.
The useful development is the reported move into an existing warehouse workflow. The sources do not disclose robot counts, parcel throughput, human intervention rates, or commercial terms. Facility-wide capacity figures and earlier demonstration results cannot establish what this installation delivers.
Highlights
CRISP: testing the contact physics behind tight assembly
Somang Lee and colleagues, Seoul National University • Submitted September 18; listed September 21 • Preprint; simulation and software
CRISP is a physics engine for tasks where many surfaces touch at once, including gears, threads, and tight-fitting pegs. It combines detailed shape representations with numerical methods for resolving contact forces.
In a simulated peg insertion with 50-micrometer clearance, CRISP completed the assembly while the tested Isaac Sim configuration failed. These comparisons depend on geometry, solver settings, and deliberately low friction. They measure simulation behavior; the paper does not demonstrate improved policy transfer to physical robots.
The public repository includes peg-insertion and gear-assembly examples, plus setup files that download a prebuilt engine. Its license permits academic and noncommercial research. That makes the release useful for examining contact behavior, with limits on commercial use and no blanket claim that it will outperform other engines across all tasks.
Read this first
Read SeeQ and watch its paired demonstrations. The same robot policy produces different outcomes when a learned evaluator chooses among its proposed moves. The remaining failures also show why a useful next step can still fall short of completing the whole job.
Read the paper