Robotics Briefing /
Robot skills get quicker to learn and use.
A robot hand starts writing after 18 seconds of calibration, and a shared exoskeleton speeds up human demonstrations. A new action model cuts the pauses between robot movements. Today’s research tackles the time it takes to teach and run useful skills, while industry announcements put the focus on factory work and delivery.
Your quick takeaways
- ETH Zurich demonstrates precise pen writing with finger motion alone. The controller learns on the physical hand, using a webcam and a laptop CPU.
- SEED-UMI collects dexterous demonstrations faster, while IMLE-VLA generates actions faster. Both report physical-robot results, with clear limits on what their headline speedups measure.
- Maven raises $100 million for industrial robotics. NVIDIA and Skild are developing contact-simulation tools, and Stellantis plans a European autonomous-delivery concept.
Papers
Rapid in-hand writing: a robot learns how its grip moves the pen
Kai Stewart, Yasunori Toshimitsu and Robert K. Katzschmann, ETH Zurich • Submitted September 10 • Preprint; physical and simulated hands
Moving a pen with fingers requires coordinating several changing contact points. This controller watches the pen through a webcam and learns how small finger commands move its tip. After roughly 18 seconds of initial motion, it follows writing paths and keeps updating that relationship.
The physical ORCA hand achieved about 0.67 millimeters of average in-plane tracking error across 16 on-paper runs. Its fingers draw each stroke; an arm repositions the hand between letters. The same controller formulation also runs on two simulated hands, and the public repository includes simplified simulation implementations.
This demonstrates a lightweight route to a specific dexterous skill without simulation training or collected writing demonstrations. The setup still uses tracking markers, a thick pen sleeve, and predefined starting grips. It controls only two dimensions, with raised, compliant paper absorbing the pen’s uncontrolled vertical drift.
SEED-UMI: human and robot share the same hand interface
Tengbo Yu and colleagues, Peking University and Delta Intelligence • Submitted September 10 • Authors list CoRL 2026; physical-robot experiments
Human fingers and robot fingers move differently, making demonstrations hard to transfer. SEED-UMI puts the same exoskeleton on both hands. Shared joint measurements and matched wrist-camera views help translate human motion into robot commands, with paired robot replay correcting errors under contact.
Across five tasks and three policy types, the method averaged 70.0% success, compared with 71.7% for training on teleoperated robot demonstrations. Each task-policy-condition combination used 20 robot trials. In a separate 30-minute AirPods insertion collection test, one operator recorded 52 successful demonstrations with SEED-UMI versus 18 through teleoperation.
The benefit is more human teaching data with similar average robot performance. That collection comparison excludes initial calibration, and changing robot hands still requires engineering. One task’s success definition also counts an initiated spray-can press without requiring actual spray, because the hand lacks enough force.
The repository currently presents project material; a complete training implementation is not established.
IMLE-VLA: fewer computation steps between robot movements
Kian Hosseinkhani and colleagues, Simon Fraser University and University of Pennsylvania • Submitted September 10 • Authors report IROS 2026 acceptance
Vision-language-action models turn camera images and instructions into robot movements. Many generate actions through repeated refinement steps. IMLE-VLA replaces that process with a single-step generator, trained to preserve multiple valid ways of completing a task.
On an NVIDIA L40S GPU, the authors measured 55 model evaluations per second, versus 15 for the original π0.5 implementation and 25 for an optimized version. In LIBERO simulation, success averaged 98.0% across 40 tasks, compared with 97.5% for π0.5. Four physical Franka Panda tasks, with 20 trials each, also showed improved success and smoother movement.
Reducing computation delays can make robot motion more responsive. The reported 3.9–6.6× reduction on hardware measures time spent running the model, excluding robot movement and other overhead. Hardware tests also use different numbers of actions between observations, so the comparison combines model and execution changes. The project’s code button currently provides no implementation link.
News
Maven Robotics raises $100 million for industrial work
Announced September 10 • Series A and company deployment claims
Maven announced a $100 million Series A led by RoboStrategy. Its initial workflows include mixed-case palletizing and tote handling, with expansion into more complex handling and assembly. The company says robot fleets already work multiple daily shifts at an unnamed Fortune 250 consumer-goods customer.
That puts the focus on sustained production use. Maven targets 100,000 autonomous operating hours by year-end and one million by the end of 2027; those remain forecasts. The announcement gives no fleet size, intervention rate, or detailed reliability data to evaluate the claimed deployment.
NVIDIA and Skild work on better contact simulation
Technical update September 10 • Company research and planned software release
NVIDIA’s update describes joint work with Skild on GPU-accelerated solvers for simulating how robots touch and manipulate solid objects. The companies plan to make these tools available through the Newton physics engine. Better contact modeling could improve training for assembly and other tasks where small physical errors cause failures.
The post also explains Skild’s previously introduced S1 model, which uses a demonstration video as a task prompt without updating model weights. Its reported 66% success measures individual steps in company tests. That figure does not establish reliable completion of entire long tasks, and the new solvers remain forthcoming.
Stellantis and UQI plan an autonomous-delivery concept for Europe
Announced September 11 • Planned collaboration and proof of concept
Stellantis Pro One and UQI Robotics announced plans to develop Box-on-Wheels, an autonomous logistics vehicle. They plan a public debut on September 14 at IAA Transportation in Hanover, followed by work to assess European customer needs and operating performance.
The partnership includes fleet integration, connectivity, and service support, all necessary for delivery operations beyond a vehicle demonstration. This remains a proof of concept. The announcement establishes neither a commercial launch date nor measured delivery reliability.
Highlights
SwarmNxt makes aerial-swarm experiments easier to reproduce
Charbel Toumieh and colleagues • Submitted September 10 • Preprint; public software and hardware setup guides
SwarmNxt combines drone setup, fleet-management scripts, and onboard navigation into a documented research platform. The public repository includes ROS packages and automation scripts, with guides covering assembly through flight. Shared infrastructure can reduce the engineering needed to test coordinated flying robots.
The authors tested six drones in open space and four with vision-based obstacle avoidance indoors. They report no collisions during roughly two hours and 30 minutes of accumulated flight, respectively. External motion capture supplies position, and depth estimation consumes about 95% of the onboard GPU, leaving outdoor independence and extra perception workloads unresolved.
Read this first
Read the in-hand writing paper, then explore its linked browser demo. The physical results and explicit setup constraints show exactly how much dexterity a small, continuously adapting controller can achieve.
Read the paper