Robotics Briefing /
Better demonstrations. More capable robot hands.
A robot can struggle before training even starts: the human teaching it may find the task hard to demonstrate. Today’s papers tackle that problem, predict what fingers will feel, and transfer simulated assembly skills to physical hardware. Industry updates show where those skills could meet factory and retail work.
Your quick takeaways
- GLIDE raised successful human-operated demonstrations from 0–1 out of 10 to 7–9 out of 10 across three tasks. Autonomous execution remained less reliable.
- DexTacWAM completed 55 of 80 trials across four manipulation tasks, versus 24 for its strongest comparison method. It learns to predict contact as well as images.
- Boston Dynamics announced an operational Atlas training center at Hyundai’s Georgia plant. Simbe reported more than 3,000 autonomous units under contract; that figure does not establish an active fleet count.
Papers
GLIDE: helping people demonstrate difficult robot tasks
Yuchen Song, Aditya Mittal and Unnat Jain, UC Irvine • Submitted September 21; listed September 22 • Authors report CoRL 2026 acceptance
Moving a plate of tomatoes with two robot arms takes careful coordination. GLIDE uses a coding agent to generate movement constraints from a task description and teleoperation code. It reviews recorded attempts and revises those constraints, helping the operator keep the plate level or release a marker gently.
Across plate transfer, marker handover and wine serving, ordinary VR control produced zero, one and zero successful demonstrations out of ten. The refined filters raised those results to seven, nine and nine. Expert-written constraints reached two, six and six.
Robot policies trained on mixed-quality demonstrations and run with the filters achieved seven, six and six successes out of ten. This separates better teaching from dependable autonomous performance.
The work shows how assistance during data collection can make difficult skills learnable. Each condition has only ten trials, and the filters use the robot’s internal state at runtime. They cannot directly see slipping objects or changing liquid levels.
DexTacWAM: predicting what robot fingers will feel
Haoran Yuan and colleagues, UIUC, UC Berkeley and Northwestern • Submitted September 21; listed September 22 • Preprint; physical robot tests
Cameras cannot reveal every change in a grasp. DexTacWAM extends a video prediction model with fingertip touch signals, then uses its predictions to choose actions. The aim is to anticipate how contact changes while a robot transfers objects, uses tongs or unscrews a cap.
Across four tasks scored as complete successes or failures, it completed 55 of 80 physical trials. Reactive Diffusion Policy, the strongest comparison method, completed 24. Each task used twenty trials per method.
The full six-task average was 70.6 points versus 38.0. Two tasks award partial credit, so that average should not be read as a completion rate. Training used about four hours of tactile recordings plus roughly 100 demonstrations per task.
The results support predicting contact alongside visible motion. Testing covered one hand design and sensor setup, and successful demonstrations supplied the task training data. Recovery from mistakes and transfer to other touch sensors remain open questions.
InsertAnything: simulated practice guides real assembly
Zhenghua Ma and colleagues, Chinese Academy of Sciences, UCAS and PaXini AI • Submitted September 21; listed September 22 • Preprint; physical robot tests
A slightly misplaced socket can make a robot jam a connector. InsertAnything trains a controller entirely in simulation, combining a target position with fingertip force measurements. On hardware, contact guides small corrections without further policy training.
In 45 physical hexagonal insertion trials with target-position errors of up to three millimeters, force feedback increased successes from 31 to 38. Average peak contact force fell from 3.08 to 2.65 newtons. Separate physical tests included nominal clearances down to 0.02 millimeters.
The released repository includes simulation environments, training scripts and real-robot deployment utilities, allowing others to inspect how the contact corrections work.
The tested skill starts with a held object and a calibrated target pose. Autonomous grasping, target localization and failure recovery still need integration. The paper’s separate 20/20 ManipulationNet result used five geometry-specific policies and a human-in-the-loop protocol, with the operator reporting task status.
News
Boston Dynamics brings Atlas training into Hyundai’s factory
Boston Dynamics • September 21 announcement • Company-reported training facility
Boston Dynamics announced that the first phase of its Robotics Metaplant Application Center is operational at Hyundai’s Georgia campus. Atlas robots are training on manufacturing work, including preparing and ordering automotive parts for assembly.
The practical value is access to real tasks and working conditions while the robot learns. The company plans a move to a facility roughly ten times larger in 2027 and further training with customers in other industries.
These are training and expansion milestones. The announcement provides no measured production throughput or intervention rate, so it does not yet establish dependable manufacturing performance.
Simbe reports more than 3,000 units under contract
Simbe • September 21 • Company-reported commercial milestone
Simbe says it has passed 3,000 autonomous units under contract. Its retail platform combines shelf-scanning robots with cameras, RFID and other sensors to track stock, product locations and pricing.
For shoppers and store teams, the value comes from finding missing or misplaced products sooner. The milestone points to demand for robots with a specific, repeatable job.
Under contract is different from installed and operating. Simbe’s claim to the largest committed fleet comes from its own research, and the release does not provide a fleet-wide uptime or intervention measure.
AMD previews onboard robotics at ROSCon
AMD • Announced September 21; event September 22–24 • Planned demonstrations
AMD’s ROSCon lineup includes Foundation Robotics’ Phantom MKI running full-body control on a Ryzen AI Embedded processor, plus navigation and simulation demonstrations.
Running control onboard matters when robots must respond without waiting for a remote server. The announcement outlines demonstrations, with no comparative performance measurements. Treat it as a preview of hardware and software integration.
Highlights
H2RBench: checking whether more human video helps
Chuyang Xiao and colleagues, CMU, University of Michigan and Toyota Research Institute • Submitted September 21; listed September 22 • CoRL 2026 paper
H2RBench compares four approaches to learning robot manipulation from human demonstrations across four tabletop tasks. More human data helped some methods and tasks, while precise insertion and longer sequences showed less consistent gains.
Across sixteen method–task combinations, simulated and physical success rates had a correlation of 0.89 under the maximum-human-data setting. That supports simulation as a way to compare approaches before costly hardware trials.
The study trained separate policies for simulated and physical evaluation. Its correlation therefore does not show that the same policy transfers directly between them. The benchmark also uses controlled camera views and mostly rigid tabletop objects, leaving everyday video and more complex handling outside its scope.
Read this first
Read InsertAnything for a concrete link between simulated learning and physical performance. Its force-feedback comparison shows how a robot can correct a misplaced target while reducing contact loads, and the released code makes the setup easier to inspect.
Read the paper