Robotics Briefing /
Better touch, clearer failure signals, and more adaptable movement.
Robots need to recognize what contact means, notice when work has gone off course, and choose movements that fit their surroundings. Three new papers make progress on those problems. On the industry side, Formic’s funding and a robotics integration update show how much of deployment still depends on financing, maintenance and connecting the pieces.
Your quick takeaways
- OpenTac improved physical insertion tasks by adding touch that senses sideways forces as well as pressure. Its small, controlled tests leave durability and broader reliability open.
- Progress Probes detect stalled or failing tasks using signals already inside a robot model. Detection is useful, but it does not by itself supply a recovery plan.
- A humanoid learned to plan body contacts across varied terrain. Meanwhile, Formic raised $50 million to expand its manufacturing robotics business, including general-purpose systems.
Papers
OpenTac: feeling the difference between alignment and a jam
Ariel Slepyan and colleagues • October 22 preprint
A plug pressing against the outside of a socket can produce pressure without being correctly aligned. OpenTac measures both pressure and shear—the sideways forces that help distinguish those situations. Its ecosystem connects customizable flexible sensor layouts, electronics, software and simulation.
On a physical SO-101 arm, adding OpenTac’s 3D tactile observations to joint-angle inputs raised peg-in-hole success from 66% to 82% and plug-insertion success from 54% to 70%. Each condition used 100 training demonstrations and 50 evaluation trials. These policies did not use vision.
The comparison supports a concrete role for touch when the important information is at the contact surface. However, the sensor’s raised bumps also change friction and contact mechanics; the researchers trained separate joint-angle-only baselines to help account for that difference.
The flexible circuit boards do not stretch. The sensing material also exhibits drift, temperature dependence and hysteresis, meaning its response depends partly on previous loading. These controlled insertion results do not establish long-term reliability across industrial tasks.
Progress Probes: noticing when a robot stops making progress
Atiksh Bhardwaj and colleagues • October 22 preprint
A vision-language-action model converts images and instructions into robot movements. Progress Probes, or P², reads that model’s internal activity with a small linear predictor trained only on successful demonstrations. It estimates task progress and flags execution that falls behind the expected pattern.
The researchers evaluated the method on six simulated tasks and six physical manipulation tasks using π0.5. The hardware evaluation contained 120 rollouts, evenly divided between successes and failures. The reported hardware AUROC was 0.78: a measure of how well scores separate success from failure across thresholds, not a claim that 78% of failures are caught.
This could make monitoring less computationally demanding than running a separate large model to watch the robot. The study evaluates detection, however, rather than a complete system that stops safely, requests help or repairs the mistake.
Subtle failures remain difficult. Spilled coffee beans may occupy too little of the image to change internal signals clearly. The experiments also focus on one model family, so performance across other robot models remains unproven.
Humanoid mobility: plan the contacts, then find the movement
Junzhe He and colleagues, ETH Zurich • October 22 preprint
Climbing a box or ladder requires more than choosing the next footstep. This framework predicts where different body parts should contact the environment, then uses a learned controller to turn those contact plans into movement. Human motion supplies training guidance; reinforcement learning expands the range of terrain the controller can handle.
The researchers demonstrate a physical Unitree G1 climbing stairs, ladders and boxes, traversing sparse terrain and recovering from some falls or failed attempts. A single policy uses onboard observations and user-specified goals without receiving a reference motion during deployment. Separate simulation experiments analyze generalization and design choices.
The useful idea is that the same contact goal can permit several workable body motions. That gives the robot room to adapt instead of reproducing a fixed trajectory.
The system handles local movement, not autonomous navigation through an entire building. Its terrain-focused representation also limits overhead contacts, hanging and manipulation. Demonstrated recovery should not be read as a measured guarantee that falls can always be handled safely.
News
Formic raises $50 million for the work of getting robots into production
October 22 announcement • Company-reported financing and operating figures
Formic announced $50 million in growth capital, combining a Series B led by Mercato Partners with debt financing from S2G Investments. Its recurring-payment model bundles robotic equipment, installation, maintenance, financing and performance monitoring.
The company reports more than 750,000 live production hours across over 100 facilities. It plans to extend its task-specific automation business into general-purpose robotics and is working with Physical Intelligence on manufacturing applications.
The significance is practical: access to a capable robot is only one part of adopting automation. Someone must integrate it, keep it running and carry the financial risk while the customer measures useful output.
Those operating figures are company claims about the existing business. They do not establish equivalent reliability for the newer general-purpose systems. The financing supports an expansion plan; it is not evidence that those future deployments already work at scale.
RoboWIZ brings perception, motion and safety partners into the same framework
NEXCOM • October 22 update from GTC Berlin, October 20–22
NEXCOM’s update describes how RoboWIZ connects robot controllers with cameras, drives, actuators, software and engineering services. The featured partners include RealSense, Stereolabs, Synapticon and Reynolds & Moore, spanning depth perception, motion control and functional-safety engineering.
This is an expansion of the public picture of a platform first introduced in August, not a brand-new October launch. Its relevance is the integration burden: AI computing must operate alongside predictable motion control and safety functions.
The announcement describes available components and partnerships but supplies no comparative deployment-time or robot task-success measurements. A platform that includes safety technology does not automatically make every assembled robot safe for every application.
Highlights
OpenTac’s hardware and software are available to inspect
Companion resource to the October 22 paper • Public repository
The OpenTac repository contains sensor-design tools, firmware, readout-board fabrication files, bills of materials and simulation examples. Its documentation connects a sensor outline to manufacturing files and then to live readings.
This makes the release more useful than a demonstration alone: the implementation is available for inspection and reproduction. Some repository tests require data that is not included, and the paper’s transfer evaluation concerns normal-contact shape rather than demonstrated transfer of a complete robot policy from simulation to hardware.
Read this first
Read the physical-experiment section of OpenTac. The plug-insertion comparison makes the value of sideways force sensing concrete, while the sensor-specific baselines show why changes to hardware and changes to learning are difficult to separate.
Read the paper