Robotics Briefing /
Learning from corrections. Moving beyond familiar conditions.
A human correction can teach a robot more than a movement to copy. Today’s papers explore how to use that feedback, fit unfamiliar parts together, and cross from land into water with one enclosed mechanism. A new benchmark also tests how closely simulated performance follows what a physical robot can do.
Your quick takeaways
- BEE reached 85% success on physical snack hanging, versus 21.7% for a comparison method using the same robot-data budget. Human takeover during training also decreased.
- InsertionWM completed 56% of simulated insertions involving ten unfamiliar assemblies after training on ninety others. Physical deployment remains untested.
- Duke’s MARBLE uses moving internal weights to roll on land and propel itself across water. The demonstrated robot follows operator commands; autonomous navigation remains future work.
Papers
BEE: learning which parts of a human correction matter
Weihui Zhao and colleagues, South China University of Technology and AgiBot • Submitted September 23; listed September 24 • Preprint; physical robot tests
When a person corrects a robot, some parts of the adjustment are consistent while others vary. BEE learns that difference, placing stronger constraints on movements people correct reliably and allowing more exploration elsewhere.
It adds a small correction policy to a frozen vision-language-action model, which converts visual observations and instructions into actions. Human interventions guide reinforcement learning while the main model stays unchanged.
On physical snack hanging, BEE achieved 85.0% success against 21.7% for RLT, a comparison method trained with human intervention. Results average three rounds of twenty evaluation trials. Both methods received the same robot-data budget and intervention protocol.
During training, humans took over 17.0% of control steps for BEE, compared with 62.1% for RLT. Evaluation ran without interventions. That suggests better use of expert help can improve autonomy while reducing supervision during learning.
Evidence remains limited to three physical tasks. Phone-charging and cloth-alignment results measure only their precision-critical phases, and the paper’s overall average also includes a simulated task. Those figures should not be treated as broad, end-to-end reliability.
InsertionWM: anticipating how unfamiliar parts will fit
Nicklas Hansen and colleagues, NVIDIA, UC San Diego and USC • Submitted September 23; listed September 24 • Authors identify IROS 2026; simulation only
Assembly robots often need a separate controller for each pair of parts. InsertionWM learns a shared world model: an internal predictor of how an action changes the situation. It combines depth-camera observations with robot state and uses predicted outcomes to plan movements.
After training on ninety assemblies, it achieved 56% success on ten held-out assemblies, evaluated over one hundred simulated trials per assembly. Those parts were unseen during training but came from the same dataset distribution.
The paper compares this with a published 7% AutoMate generalist result. The exact AutoMate training and test split is unknown, so the comparison is not a matched test on identical held-out parts.
The promise is less part-specific engineering when a production line changes what it assembles. That remains a research direction: all experiments were simulated, and the setup begins with a grasped part and supplies estimated socket poses. Camera differences, contact physics and physical deployment still need testing.
MARBLE: a rolling robot keeps its moving parts inside
Niko Weaver and colleagues, Duke University • Submitted September 23; listed September 24 • Preprint; physical locomotion demonstrations
Exposed wheels, joints and propellers can catch debris or admit water. MARBLE encloses its motors inside a sphere roughly thirty-nine centimeters across. Three sliding internal weights shift its center of mass, making the shell rotate.
On land, that rotation rolls the robot. On water, small passive fins attached to the shell provide propulsion. Physical demonstrations show land travel, movement across the water surface and a continuous land-to-water transition without changing propulsion hardware.
In separate fourteen-second recordings, reported mean speeds were about 0.75 meters per second on land and 0.38 on water. A learned controller trained in simulation also ran on the physical robot without further hardware training.
The design shows how one enclosed mechanism can serve two environments. These are demonstrations under operator commands, with motion measured externally from video. Autonomous navigation, submerged operation and long-duration field reliability remain unestablished. The authors promise a hardware and software release; its availability was not verified.
News
Xcanbot brings seated autonomous mobility to REHACARE
Xcanbot • September 23 announcement; September 23–26 event • Company-described demonstration
Xcanbot announced the trade-sector debut of its Mate X seated mobility robot at REHACARE in Düsseldorf. The company says riders select a destination and the machine navigates using lidar, distance sensors and cameras.
For people whose daily journeys are limited by walking distance, this points toward mobility assistance with less continuous steering. The current milestone is a demonstration and distribution push, with a European crowdfunding campaign still in preparation.
The announcement supplies no independent navigation benchmark, intervention rate or evidence of broad customer deployment. It describes a consumer product, so the trade-show appearance should not be read as medical approval or demonstrated suitability for every rider.
Epson adds its first collaborative robot
Epson • September 22 • Catch-up: product launch and announced availability
Epson introduced the AX6, a six-axis arm with power- and force-limiting features, no-code programming, Python support and a simulator for testing programs before running them.
The addition gives manufacturers another option for applications where people work near robots and equipment needs to be redeployed. Epson says the AX6 is available through its distributor network, with an ISO Class 5 cleanroom rating.
Operation without traditional barriers still depends on an application-specific risk assessment. The announcement provides product capabilities rather than measured improvements in factory throughput or installation time.
Highlights
X2Real: checking whether simulation predicts physical performance
Lian Ruan and the X Square Robot team • Submitted September 23; listed September 24 • Benchmark preprint; paired simulation and hardware evaluation
X2Real introduces forty-four tasks spanning ten capability dimensions, with separate training and evaluation conditions. Its aim is to make simulated tests more useful for judging general-purpose manipulation.
In a paired check using one robot policy across eight tasks, simulated and physical completion rates had a correlation of 0.74. The higher 0.84 figure measures partial task progress. Precision tasks exposed substantial gaps: placing glasses on a shelf succeeded 70% of the time physically and never in simulation.
That makes the benchmark useful for studying where simulation misleads, alongside where it agrees. The alignment check covers one policy and a small task set; soft materials and tactile sensing remain outside the benchmark’s current scope. The repository linked in the paper returned unavailable during review, so a usable code release is not yet verified.
Read this first
Read BEE for a concrete example of making human assistance count. Its physical snack-hanging comparison measures both task success and the amount of human takeover during training, connecting learning quality with the effort required to teach a robot.
Read the paper