Robots learn to work with moving targets.

EgoThumb: an extra digit that follows the hand

Jianyu Hou and colleagues, UC San Diego and Sudo AI • October 27 arXiv release; IROS accepted • Catch-up

EgoThumb helps one hand hold an object while its biological fingers do something else, such as unscrewing a bottle cap. A camera facing the palm observes the interaction, and a controller learned from demonstrations decides how to move the additional thumb. Motors sit on the forearm; inference runs on an external computer.

After fifty demonstrations per task, average success across four physical tasks was 90% for the demonstrator and 92.5% and 87.5% for two additional users. Each user completed ten evaluation attempts per task.

The interesting possibility is assistance that responds to context without requiring a separate command for every grasp. However, this small study does not establish everyday reliability. Expanding the initial and destination areas dropped average ball-transfer success to 17%. Transferring across people proved easier than transferring across task layouts.

• Paper ↗
• Project and hardware ↗

A drone learns when to pass through a swinging opening

Riccardo Bianco and colleagues, University of Zurich • October 27 arXiv release; CoRL 2026 listed • Catch-up

Passing through a moving gap requires arriving at the right moment with the right position and attitude. This work trains a controller in differentiable simulation: the simulator can calculate how changes to earlier actions would improve the eventual trajectory, giving learning more guidance than a final success-or-failure score.

The vision-based controller achieved 99.3% success in its training simulation and eight successes in ten scored physical flights. It used onboard visual features and inertial measurements, without further training on the real drone.

The result demonstrates a controlled transfer from simulation to hardware. It also shows how much conditions matter: the physical opening had LED-marked corners and swung in one plane. The system uses extracted geometry rather than unrestricted camera images. Unmarked openings, unpredictable motion and extended loss of visual contact remain untested, and ten flights leave considerable statistical uncertainty.

• Paper ↗

StabBot: meet a suspended pipe at its own speed

Mohamed Ashraf Mostafa, Taher El-Fakharany and Samir Khaled • October 27 preprint • Simulation-only catch-up

Guiding a hanging pipe into another pipe is difficult when the load is swinging. StabBot predicts that swing from stereo-camera measurements, moves a roller-equipped guide alongside it, and uses contact feedback for final alignment. A learned correction refines the underlying controller.

In 200 paired simulated jobs spanning nine pipe sizes, StabBot seated every joint without unwanted contact, compared with 74% for a fixed guide. Median contact speed fell from 18.4 to 2.3 centimeters per second. Prediction and speed matching already achieved full success; the learned correction further reduced forces.

The useful idea is to reduce relative motion before making contact with a heavy load. All reported results, however, come from a quarter-scale MuJoCo model expressed at full-scale equivalents. Real camera images, physical pipe tests and rig-specific effects are future work. Code and data are promised on acceptance, not established public releases.

• Paper ↗

Otto Group and Sereact plan robots for unpredictable returns

October 28 announcement • Planned staged deployment

The partners plan to introduce ten dual-arm systems at an Otto Group logistics center in Poland, with expansion toward 25 dependent on initial results. Broader use across the logistics network would follow a successful second phase.

Returns are a demanding test because items arrive in changing conditions and packaging. Sereact says its Cortex model can grasp, assess and route unfamiliar items. Its training pipeline uses NVIDIA simulation and world-model tools to vary the situations represented in training.

Sereact will operate and maintain the systems under a service model. People retain supervision, exception handling and final quality decisions. That division of responsibility matters: automating the grasp does not remove the need to resolve an uncertain return.

This is a deployment plan and joint development project, not evidence that 25 robots already run reliably. Throughput, error rates and the human workload around exceptions will determine whether it scales.

• Announcement ↗

Foundry Robotics raises $100 million for flexible assembly

October 28 announcement • Company-reported funding and facility expansion

Foundry Robotics announced a $100 million Series A led by Marlinspike Partners and a 130,000-square-foot facility in Emeryville. It builds and operates factories combining robotic assembly cells with manufacturing specialists, serving complex commercial and defense hardware.

The business offers manufacturing as a service: customers purchase production capacity rather than building a dedicated line. Foundry says its AI models learn assembly building blocks, while factory software coordinates the robots and production process. Funding will support the facility buildout, model deployment and hiring.

The robotics question is whether flexible cells can change products economically while maintaining quality. Owning the factory gives Foundry control over both the machines and the conditions they work in, which could make integration easier.

Funding and floor space establish resources, not manufacturing performance. The announcement provides no independently verified comparison of throughput, defect rates or human interventions across products. Those measurements will be needed to assess the claimed flexibility.

• Announcement ↗