Robotics Briefing /
Teaching robots what to feel, what to check, and when to stop.
A robot can copy a hand movement and miss the contact that makes it work. It can finish one step and fail to notice that the result has come undone. Today's research makes those gaps easier to see, with tactile demonstrations, explicit checks during long tasks, and a drawing benchmark that exposes weak instruction following. Outside the lab, residential parking robots face competing requests, while Agility prepares to finance its next stage.
Your quick takeaways
- TouchBridge collected human demonstrations seven times faster than teleoperation, with robot success close to—but below—the teleoperated-data comparison.
- LOGIV improved physical task completion from 12 of 30 trials to 17 of 30 by checking and repairing task dependencies; precise insertion remained difficult.
- BURGER's newly released code helps generate robot drawing data, but its real-robot deployment and policy-training code are outside this release.
Papers
TouchBridge: preserve the contact, not just the hand movement
Haozhe Zhang and colleagues · Preprint first submitted October 29; Friday-listing catch-up · Physical robot tests
Human and robot fingers have different shapes. Matching their positions does not necessarily preserve where they touch an object. TouchBridge pairs a wearable recording interface with corresponding robot sensors, then transfers demonstrations using fingertip contact patches and the spatial relationships between contacts. It also converts the imagery into robot-like observations for training.
The team fine-tuned a pretrained vision-language-action model—a system that turns images and instructions into movements—using 60 demonstrations per task. Across four physical tasks, including turning a book page and screwing in a bulb, converted human data yielded 72.5% average success versus 77.5% for an equal amount of teleoperated robot data. Each policy received 20 trials per task.
The collection advantage is substantial, but the evidence is narrow: four tasks on one robot setup. Touch measurements guide demonstration conversion; they are not inputs to the deployed policy. This supports a faster route to useful training examples, not a claim that the robot has learned general touch-based control.
LOGIV: a finished step can still need repairing
Enqiao Lu and colleagues · Preprint first submitted October 29; Friday-listing catch-up · Simulation and physical tests
Placing a support block does not guarantee it stays in place. LOGIV represents a task as connected actions and conditions, distinguishing whether an action finished from whether its result still supports the remaining plan. Visual observations check those conditions; after a failure, the system repairs the affected part while retaining usable work.
It sits above an existing robot policy without adding trainable components. Experiments covered 20 simulated dual-arm tasks, with separate evaluations using predefined disturbances. Physical tests covered tomato placement, block assembly and cable insertion on two robot platforms. Those trials used no injected disturbances.
The small physical study suggests that checking dependencies helps, but it does not establish a precise effect size. Cable insertion improved only from two successes to three in ten attempts: repairing the plan could not reliably make the plug seat. The distinction matters because better reasoning can recover a sequence without solving the contact precision required by its hardest step.
BURGER: a blank canvas tests whether instructions matter
Xiangyu Xu and colleagues, Tsinghua University and Shanghai Qizhi Institute · Preprint October 29; code release October 30
Every drawing task begins with the same blank canvas, so a robot cannot infer the requested object from the starting scene. BURGER uses this simple setup to test vision-language-action models with 28,200 simulated demonstrations across 282 tasks and five model variants.
Some models improved more consistently as data grew and could combine familiar objects in unseen spatial arrangements. Unseen object categories remained a problem. Other failures included drifting from the drawing area or continuing to draw after finishing an object—useful distinctions that a single success score can hide.
A separately trained Franka Panda policy used physical demonstrations and responded to human-added strokes, including continuing a bicycle after a person supplied a wheel. This is not evidence of direct simulation-to-real transfer. The benchmark also simplifies contact to line drawing with a rigid tool, so its results cannot establish general manipulation ability.
News
Hyundai tests parking robots against everyday apartment requests
Hyundai Motor Group · Announcement October 30; demonstration October 29 · Residential pilot
Hyundai Motor, Hyundai WIA and Hyundai Engineering & Construction demonstrated parking robots at Hillstate The Wave City in Siheung, South Korea, in a residential parking area with more than 50 spaces. Scenarios covered simultaneous arrival and departure requests, moving double-parked vehicles, and temporarily retrieving a car before returning it.
Residents could request retrieval through an apartment wall panel. Hyundai reports that the demonstrated retrieval took about three minutes. Connecting the robot to building barriers, resident controls and vehicle inspection systems is the practical advance: moving a car is only one part of delivering a parking service.
This remains a pilot. The announcement supplies no long-term reliability or intervention rate. Hyundai also says its separate management system can control up to 200 robots; that is a stated capacity, not the number demonstrated at this site.
Agility expects about $465 million as its public-market transaction approaches
Agility Robotics and Churchill Capital Corp XI · October 30 · Company announcement; closing pending
Agility and Churchill XI expect approximately $465 million in gross capital associated with their proposed business combination and related financing, before transaction costs. That total includes about $164 million from the trust account, $201 million in committed private investment, and $100 million already received. It should not be read as $465 million of newly delivered cash.
The companies expect closing on November 4 and Nasdaq trading on November 5, subject to shareholder approval and other conditions. The robotics relevance is funding for commercial expansion and support as Agility prepares Digit 5 deliveries. Early-access deliveries remain targeted for the first half of 2027.
Capital can support manufacturing and customer service, but it does not demonstrate robot uptime or economic performance. The transaction is still pending, and the delivery dates remain company targets.
Highlights
Available now: BURGER's data-generation pipeline
Open-source release · October 30
The public repository contains code for turning sketches into ordered strokes, recording simulated robot demonstrations in ManiSkill, LIBERO or Isaac Sim, and exporting datasets. That makes the data-creation method inspectable. The release explicitly excludes prompt-to-image generation, the paper's exact evaluation splits, policy training and evaluation, and real-robot deployment. Those boundaries matter when judging how much of a research result a code release lets others reproduce.
Read this first
Read LOGIV's physical-results section. Its cable-insertion failures show exactly where repairing a task plan stops helping and reliable contact control has to take over.
Read the paper