Digit 5, Navy contracts, tactile data

Agility has put numbers on its humanoid business case, Gecko quadrupled a Navy contract, and Daimon is trying to make touch a first class training signal. Also today: red teaming for physical AI, a muscle powered swimmer, and a Swift client for ROS 2.

Digit 5 bets on deleting the safety fence

Agility Robotics unveiled Digit 5 on September 15 and published the specification that matters to anyone designing a humanoid cell (Agility, product page). It lifts 50 lb (22.7 kg) repeatedly, 40 percent more than Digit 4, on a new leg design built around cycloidal actuators. It stands 1.81 m, weighs 129 kg, reaches 2.2 m (Digit 4 managed 5.5 ft), and takes swappable end effectors on ISO standard flanges.

The duty cycle is the commercial claim: 90 minutes of runtime on a 9 minute charge, a 10:1 run to charge ratio against 2:1 on Digit 4, which Agility says allows more than 20 hours of work in a 24 hour day (Nyobolt release). Fast charging instead of battery swapping moves the constraint into the building: chargers, grid capacity and thermal management decide whether the ratio survives a real warehouse.

Safety is the consequential part: human detection across several sensor types, visual and auditory cues that signal motion intent, and an independent safety controller that stops, avoids or seats the robot when a person is too close. Agility says this removes the physical barriers conventional automation needs. Digit 5 is a launch partner for NVIDIA’s Halos for Robotics platform on IGX Thor, and Agility contributes to ANSI/A3 TR R15.108 and ISO 25785-1, the draft humanoid safety standard. It also says Digit was the first humanoid to pass an OSHA administered field evaluation on a customer line, a claim to treat as the company’s framing until the standards publish.

Co-founder Jonathan Hurst supplies the counterweight in IEEE Spectrum: Digit 5 is “heavy and bulky, lacking the graceful natural motion of its predecessors,” and further from the robot he set out to build than he expected. The machine was optimized for payload, uptime and certifiable safety, not for the dynamic locomotion research that produced Cassie. Agility cites more than $300 million in multi year orders and 65,000 hours of operational data, and its RoboFab plant in Salem, Oregon covers 70,000 square feet with capacity above 10,000 units a year.

Gecko turns inspection into a program of record

The U.S. Navy and GSA quadrupled the ceiling on Gecko Robotics’ contract to $300 million, six months after a $71 million IDIQ award (Gecko, The Robot Report). Gecko says the robots save up to $8 million in repair costs per ship and find structural issues up to 50 times faster than manual methods. Those are vendor figures, not audited results.

Komodo scans through thick non-skid deck coatings without stripping them and produces a full deck heatmap of corrosion and material thickness, so defects are scoped before a ship enters drydock instead of being found after work starts. Wall climbing robots, drones and fixed sensors feed Cantilever, Gecko’s data platform, which models hulls, decks, welds and components. The hard engineering is registration and repeatability: tracking corrosion over years only works if successive scans align to one reference frame with calibrated measurements.

Daimon wants touch measured at 30 Hz

Daimon Robotics sells infrastructure rather than a model: tactile sensors, a handheld capture device called Umi, simulation, and the models trained on the output (The Robot Report). Tactile feedback runs at 30 Hz. The company says it has discussed more than 10,000 hours of collected data, with the longer term Daimon-Infinity effort targeting over 1 million hours.

Its Dr. Zhu is blunt about simulation: “not good enough. It’s kind of terrible.” Daimon calibrates simulated friction and contact behavior against real measurements before moving data into sim. At IROS 2026 it demonstrated threading bracelets, applying heat transfer designs to tote bags and clearing broken glass, using foundation models trained across scenarios rather than one policy per task. Daimon-TWM is its tactile grounded world model.

Vision locates a part. It does not tell you whether the part is seated, slipping or crushed. Assembly failures are usually contact failures, and the field still lacks a shared schema and benchmark for tactile policies. Watch whether Daimon publishes those, or only demos. Same argument as the robot data refinery piece: the pipeline is the asset.

Laplace sells failure before deployment

Laplace Intelligence, an Ann Arbor spinout from the University of Michigan founded a year ago, red teams robots and autonomous vehicles in simulation (TechCrunch). Its environment is a post-trained version of NVIDIA’s Cosmos, grounded in map data and imagery, with AI agents that design corner cases from collision records and news instead of waiting for a human to invent them.

The detail that matters is interactivity: other actors react. A test car swerving into oncoming traffic meets a car that avoids it rather than driving through it, a known failure of purely generative world models. CEO Howie Sun says customers include the trucking firm Bot Auto and a major Japanese automaker, and that the company is generating revenue and a modest profit. As humanoids leave structured cells, nobody has a catalogue of edge cases for a home. The open question is coverage: an adversary can generate failures forever, and insurers will want evidence that the tested distribution is representative. See the world action models piece.

A muscle powered swimmer the size of a stick of gum

MIT’s Raman Lab built a swimming robot from a gel film the length and width of a stick of gum, with a layer of light sensitive muscle cells thinner than a human hair on each fin (MIT News via Robohub, paper in Advanced Functional Materials). Flashing light on one fin makes it flap, so the illumination pattern sets direction and speed.

It covers about four body lengths per minute at top speed and navigated a simple watery maze. First author Maheera Bawa and Ritu Raman call it the first very thin, two dimensional, muscle powered robot capable of locomotion, built from far fewer cells than the bulky 3D constructs the field has used. It is a data point for soft actuation: living muscle brings compliance and self repair that no servo matches, at the cost of keeping cells alive.

Swift joins the ROS 2 graph without rcl

The native Swift client library for ROS 2 reached 3.0.0, split into three packages that mirror upstream: swift-ros2-rosidl 1.0.0 with the XCDR v1 codec and message, service and action protocols, swift-ros2 3.0.0 with the runtime, and swift-ros2-common-interfaces 1.0.0 with generated standard messages including tf2_msgs (announcement, GitHub). It speaks Zenoh, round trip compatible with rmw_zenoh_cpp, or DDS through rmw_cyclonedds_cpp, with verification reported against Humble, Jazzy, Kilted and Lyrical. Services and actions are keyed by service level type hashes.

The practical win is an iPad or Mac UI that subscribes to sensor topics and calls actions with no Python bridge and no cross compiled rcl. The risk with any alternate client is QoS and graph semantics drifting from what rclcpp peers expect, the material in the QoS policies guide. At 28 stars, read the source before putting it on a robot.

Quick Bites

• robot-data-diff is a CPU only CLI that diffs two revisions of a LeRobot dataset, local or on the Hub, reporting episodes added, removed or modified with frame counts, action statistics and idle ratios, plus an HTML report (Discourse, GitHub, PyPI). Version 0.1.0, Apache 2.0.

• HUMXN told Tom’s Hardware it will subsidize plumbing, electrical, HVAC and carpet cleaning in Minneapolis, Chicago and Miami in exchange for recording technicians with wearable capture gear (Tom’s Hardware). Its site returned no content when checked, so treat the program as announced rather than operating.

• ROS PMC minutes for October 13 list new committer candidates, a proposed in person PMC summit for roadmapping, and buildfarm regressions including undefined DDS PubSubType references on Rolling and Lyrical traced to CycloneDDS (Discourse).