Robotics Briefing /
Robots face the limits of what they can see.
A humanoid keeps walking through partial camera failure. A ground robot borrows a drone’s view above tall grass. From failure detection to factory welding, today’s developments test how robots handle imperfect conditions.
Your quick takeaways
- CAP completed 19 of 20 physical trials with partly blocked vision. Full camera cover still defeated every tested platform and gap crossing.
- EVPeriscope gives a ground robot an aerial view. FARM detects signs of failure, but its performance drops on unfamiliar tasks.
- FANUC targets welding setup, AgiBot reports gains from more training data, and Pony.ai begins driverless passenger tests in Zagreb.
Papers
CAP: keeping a humanoid moving when vision degrades
Catch-up • Hongjin Chen and colleagues • Submitted September 10; authors report CoRL 2026 acceptance • Simulation and Unitree G1 tests
CAP combines depth-camera information with measurements of the robot’s own body and joints. A learned model cleans up corrupted depth observations, while training exposes the controller to missing and unreliable visual information. This helps it use whatever useful view remains.
The controller transferred from simulation to a physical Unitree G1. It completed 20 of 20 trials with clean vision and 19 of 20 with partial occlusion, across stairs, platforms, gaps, and mixed terrain. Each terrain-condition combination used five trials.
Full camera cover exposed the boundary: five successful stair trials, but zero successful trials on platforms, gaps, or mixed terrain. Losing sight of an edge can remove information that body sensors cannot replace. Safe recovery after a failed step remains future work; the project provides videos and labels its code as coming soon.
EVPeriscope: a drone becomes a ground robot’s lookout
Catch-up • Dexter Ong, Vijay Kumar and Pratik Chaudhari, University of Pennsylvania • Submitted September 10; project lists ISRR 2026 • Physical field experiments
Tall grass can block a ground robot’s cameras. EVPeriscope sends a small drone overhead and shares its elevated camera view with the ground robot. An upward-facing event camera tracks the rapid brightness changes from spinning propellers, helping keep the drone in position.
The drone-control system runs at 200 updates per second using onboard sensing and computation. In a physical demonstration, the pair navigated through dense foliage around a tree toward a waypoint. The public repository includes perception and control code, hardware material, and setup instructions.
This could help robots inspect vegetation or structures beyond their normal view. The demonstration covers one cooperating pair, and the detector cannot separate interference from another drone’s propellers. Localization also inherits drift from the drone’s camera-based position estimate; some separate mapping experiments used a human-operated ground robot.
FARM: reading failure signals inside a robot model
Catch-up • Haoran Pei and colleagues • Submitted September 10 • Preprint; simulation and recorded physical-robot evaluations
A robot needs to recognize when a task goes wrong before it can respond. FARM trains a small detector on internal states from a frozen world model, which predicts how a scene changes. Only the detector’s roughly 34,000 parameters learn from success and failure labels.
In a matched simulation comparison, its failure-ranking score reached 0.837 AUROC on familiar tasks and 0.657 on three unseen tasks. That metric measures separation between successful and failed runs, with 1.0 perfect and 0.5 chance. Another method performed better on the unseen tasks.
Physical-robot evaluations also show uneven transfer, with improved results after training the detector on target-robot examples. Failure scores could support stopping or requesting help, but this study does not demonstrate a recovery system. The public implementation includes detector training and evaluation; model weights, robot recordings, and world-model feature extraction remain outside the release.
News
FANUC aims to turn engineering drawings into welding programs
Catch-up • Announced September 11 • Company claims; demonstration and shipments planned
FANUC’s AI Welding Agent uses Google Cloud technology to interpret component drawings and generate welding parameters and robot motion programs. Operators can review or adjust the output before execution. The goal is to reduce the specialist programming needed when parts change.
FANUC plans a demonstration at Tokyo’s International Welding Show beginning September 16, with shipments scheduled by year-end. Its announcement provides no measured setup-time savings or weld-quality benchmark. Claims of zero setup and zero teaching still need evidence across varied parts and production conditions.
AgiBot’s GE-Act 2.0 shows what more robot data buys
Catch-up • English announcement dated September 10; newsroom lists September 11 • Technical report first submitted September 4
AgiBot’s world-action model predicts future scenes and generates movements to reach them. Its report tests models directly on 100 physical manipulation tasks across two robot designs, without extra training for each test task.
Increasing joint-training data from 300 to 30,000 hours raised average success from 17.1% to 44.1% on G1-OP and from 13.4% to 31.1% on G2-90D. Each task, robot, and data-scale combination used ten trials, with held-out objects and environments.
Those gains support investing in broader robot experience. Success still stays below half of attempts on both robots, and the comparison does not hold training computation constant. The study establishes gains within this setup; dependable general-purpose manipulation remains unproven.
Pony.ai and Verne begin driverless passenger tests in Zagreb
Catch-up • Announced September 10 • Company-reported public-road testing
The partners have started fully driverless passenger test rides along a 22-kilometer route connecting Zagreb’s airport with Verne’s headquarters and a business district. Their existing commercial service used an onboard autonomous-vehicle operator.
The change tests whether the technology and local operating team can support service without that onboard role. The earlier fleet’s reported 200,000 kilometers include operations with an operator. Those miles cannot establish the reliability of the new driverless tests, and wider commercial driverless service remains a future step.
Highlights
A portable welding arm targets construction-site setup
Catch-up • FANUC update September 11 • New system demonstration planned; robot arm previously released
FANUC’s structural-column welding system uses an 11-kilogram CRX-3iA arm mounted on a magnetic base. Sensors identify mounting orientation and weld locations, aiming to simplify moving the robot between jobs.
The company plans a September 16 demonstration and describes earlier shipyard use of the arm. The new construction system still lacks published field productivity and reliability measurements. Its practical question is how much setup and repositioning time remains around each weld.
Read this first
Read CAP, then watch its cover-and-uncover demonstration. The trial table makes a useful distinction between coping with degraded vision and handling complete sight loss on terrain where the robot must see ahead.
Read the paper