Robotics Briefing /
Robots learn to handle the world beneath them.
Rough trails, thin obstacles, buried roots, and bubbles all change what a robot needs to understand. Today’s research shows how those details shape performance, from predicting a rough ride to learning underwater manipulation. Faraday Future’s latest launch also puts the gap between product availability and dependable field performance in focus.
Your quick takeaways
- Feel-WM completed 24 of 30 off-road runs without intervention, versus 14 for its comparison system. Both used less than an hour of physical driving data for adaptation.
- A safety filter eliminated observed collisions in small humanoid obstacle tests. Transparent surfaces still caused trouble, and one demonstration used cardboard to cover glass.
- ULOHA brings two-arm robot learning underwater. Adding bubbles cut one task’s success from 10 of 10 trials to 3 of 10, showing how sensitive learned skills can be.
Papers
Feel-WM: predicting how rough ground will feel
E-In Son and colleagues, Seoul National University • Submitted September 17 • Preprint; weekend catch-up
A path can look clear while making a robot slip, tilt, or shake. Feel-WM combines camera images with measurements of the robot’s own motion. Its world model predicts the physical effects of possible routes and weighs failure risk against progress toward the goal.
On a Clearpath Husky, it completed 24 of 30 runs across root-covered, downhill, and uphill courses. NoMaD, the comparison controller, completed 14. Success required reaching the goal without operator intervention; Feel-WM also recorded a smoother ride on every course.
Both systems adapted using less than an hour of driving data from different trails. Feel-WM ran onboard the robot and replanned every 0.8 seconds, making its predictions useful within the hardware’s computing budget.
The practical lesson is that navigation benefits from predicting the robot’s response to terrain. The physical evidence covers only three courses on one wheeled platform. Tests involving a legged robot took place in simulation, and the hardware results required real driving data.
RoM-Nav: giving a humanoid an extra collision check
William D. Compton and colleagues, Caltech and Amazon Safe Autonomy Frontiers Lab • Submitted September 16 • Preprint; weekend catch-up
RoM-Nav first learns navigation with a simplified model of motion, then uses that knowledge to train a full humanoid controller. A separate safety filter adjusts movement commands when obstacles get too close.
In Unitree G1 tests with unfamiliar ladder obstacles, collisions fell from two in ten trials to zero with the filter. Against thin tubes hanging at head height, they fell from four in ten to zero. Every trial reached its goal, showing why goal completion alone can hide unsafe behavior.
Those safer routes sometimes took longer. In the ladder tests, average travel time rose from 7.5 to 11 seconds. The results support checking a learned controller’s commands before the robot executes them.
Ten trials per condition cannot establish dependable safety. Glass remained a blind spot: researchers covered a glass door and window with cardboard in one demonstration. Although the navigation policy receives no map, the test system uses mapping and relocalization to supply accurate goals.
Mechanical weeding: a robot dog moves the tool with its body
Ruben Beumer and colleagues, Eindhoven University of Technology • Submitted September 17 • Preprint; weekend catch-up
This prototype attaches a milling tool directly to Boston Dynamics Spot. With its feet planted, the robot shifts its body to position the tool over a weed. A camera estimates the target location, allowing the legs to do positioning work without an added robot arm.
Indoor tests reached 72 of 88 artificial-plant targets, about 82%. That measures targeting, not successful removal of living weeds. Outdoor tests showed real weed removal when the system correctly located the roots, but the paper does not establish an outdoor success rate.
The design illustrates a useful option for mobile manipulation: a robot’s existing body motion can position a specialized tool. It also exposes the difficulty of turning a mechanical demonstration into a working farm system.
Larger plants obscured root locations, and throughput remained far below comparable wheeled systems. The current navigation also lacks a method for avoiding steps on crops. All reported tests used remote processing, so the proposed onboard improvements still need evaluation.
News
Faraday Future launches robots for education and inspection
Faraday Future • September 19 launch and company release • Product announcement
Faraday Future announced nine device configurations across five models, alongside packages for K–12 education, research, security, and inspection. Its lineup includes the compact Master Mini humanoid, which the company lists from $9,990, and several Aegis robot configurations.
The company’s pitch combines hardware with software and support for specific jobs. For schools and facility operators, that packaging could make adoption easier than assembling separate components and building each workflow from scratch.
Pricing and product availability are useful signals, but they leave the operational questions open: how often does a robot need help, which conditions cause failures, and what does each completed job cost? Treat the launch as a company announcement. The practical value will depend on evidence from sustained customer use.
Highlights
ULOHA: learning two-handed tasks underwater
Masato Kobayashi and Takeru Tsunoori, University of Osaka and Kobe University • Submitted September 16 • Preprint; weekend catch-up
ULOHA pairs two underwater robot arms with two dry control arms for collecting human demonstrations. Cameras record the work, and learning software turns those examples into autonomous actions such as passing blocks, opening lids, and catching a rising sponge.
Using the ACT imitation-learning method, separate task-specific models completed 68 of 90 trials across nine underwater tasks. Each task received ten evaluation trials with a 60-second limit. Seven tasks used ten demonstrations each; the two floating-sponge tasks needed fifty.
Bubbles exposed an important weakness. Without retraining, one sequential-transfer task dropped from ten successes to three out of ten. A shared lifting task stayed at ten, so the disturbance affected behaviors differently.
The platform offers a way to study how water changes learned manipulation. These controlled tank experiments do not establish open-water reliability, and the bubble tests cannot separate visual interference from changes in water movement. The project page currently says more details are coming.
Read this first
Read Feel-WM for a concrete example of what a world model can add to robotics. Its field tests connect predicted slip, tilt, and vibration with a simple outcome: whether the robot finishes a rough route without help.
Read the paper