Robotics Briefing /
Faster reactions, longer plans, steadier hands.
Robots need to make decisions at several speeds: choosing a route through a long task, correcting an arm movement, and responding when their body shifts. Three new papers explore those different problems. A separate industry update asks how a robot should respond when the person giving instructions changes their mind. The evidence ranges from simplified simulations to physical manipulation trials, and those distinctions matter.
Your quick takeaways
- RTPolicy reports 72% success across three physical tasks; its embedded-board speed measurement is a separate result.
- A hierarchical world model improves long-range planning in simulation, with substantial failures still present in visual tests.
- An online controller learns how a moving robot's body affects its arm without retraining the underlying locomotion policy.
Papers
RTPolicy: faster decisions from three camera views
Xianbao Hou and colleagues · Preprint first submitted October 30 · Physical tests and simulation
A robot that takes too long to process an image may act on a situation that has already changed. RTPolicy combines information from three cameras, compresses it, then generates a short sequence of movements in four computational steps. The aim is to make frequent visual corrections affordable on smaller computers.
On two ROKAE arms, it averaged 72% success across drawer handling, stacking and assembly, compared with 52% for π0.5. Each task used 200 demonstrations and 50 test attempts. Assembly remained difficult: RTPolicy completed only 30% of those trials.
The physical success tests ran on an RTX 4090. Separately, the paper measured 30.35 milliseconds per inference—32.9 Hz—on an RDK S600 using INT16 numerical precision. That timing result does not establish the same task success on that board. The study also uses task-specific models, and some comparison methods receive only one camera view. Faster inference is useful evidence, but it is one part of a working robot system.
ID-WM: planning the intermediate goals before the movements
Gaoyue Zhou and colleagues · Manuscript first submitted October 30 · Project lists NeurIPS 2026 · Simulation
Long plans are difficult for a world model—a learned predictor of how an environment changes—because prediction errors accumulate. ID-WM divides the problem between a model that proposes reachable intermediate states and a lower-level model that finds actions to reach them. This lets it combine short pieces of experience into longer plans.
In visual PushT, a simulated task involving a T-shaped block, ID-WM succeeded on 50% of goals 75 steps away. The single-level DINO-WM comparison scored 0%, while another hierarchical model scored 14%. ID-WM used DINO-WM underneath, making this a useful comparison of what the extra planning layer contributes.
The harder random-goal setting still reached only 26% success. These are simulation results, not physical manipulation trials. The authors also identify expensive search at inference time as a limitation. The work shows a promising way to extend planning distance; it does not yet resolve the speed and reliability demands of an everyday robot.
Learning a moving body's effect on the hand
Tianlin Zhang and colleagues · Preprint first submitted October 30 · Simulation and physical tests
An arm mounted on a walking robot cannot assume its base stays still. A locomotion controller may react differently to similar commands as balance and contact conditions change. This study keeps that controller fixed and learns its resulting body motion online, meaning the model updates as the robot moves.
The method separates more predictable responses from disruptive ones and reduces the influence of unreliable observations on learning. A predictive controller uses the resulting estimates to coordinate the base and arm. Physical tests used an M1 wheel-legged platform with a Kinova arm, including grasping while moving and opening a door.
In the hardware trajectory-tracking example, mean hand-position error was about two centimetres during the final 70% of the run, after larger initial errors. The moving-grasp example finished below one centimetre, but errors rose during transitions. These traces show adaptation over time, not guaranteed centimetre accuracy throughout a task. Testing different arm types and mounting positions was conducted in simulation; the hardware evidence uses one platform configuration.
News
Ludi 0.2 focuses on understanding the person nearby
KRAFTON / Ludo Robotics · Catch-up · October 30 announcement and Berkeley presentation
Ludo Robotics presented Ludi 0.2 at the Bay Area Robotics Symposium, describing improvements in interpreting human intent, remembering context, recognizing gestures and coordinating navigation with manipulation. Its earlier system combined a vision-language model for dialogue and reasoning with specialized robot capabilities.
The practical problem is familiar: instructions can be ambiguous, preferences change, and gestures carry information that words omit. A robot needs to decide when to act, when to ask for clarification and how to adjust without losing the task's context.
The announcement describes company-reported capabilities but supplies no numerical Ludi 0.2 evaluation results or reproducible comparison. It should be read as a research update, not evidence of dependable home deployment. The proposed Ludi 1.0 foundation model remains in development. Distinguishing that future unified model from the current system matters when assessing what has actually been demonstrated.
Highlights
The release details behind the results
Resources accompanying the October 30 papers
RTPolicy's public repository contains model code, training and export tools, configurations and a RoboTwin adapter. It also lists model checkpoint packages. Access to the implementation makes the method easier to inspect, though the existence of code and download links is distinct from independent reproduction of the reported scores.
ID-WM's project provides explanatory videos, while its code is still marked as coming soon. Both resources help explain the methods, but they offer different levels of access for checking the results.
Read this first
Read ID-WM's visual-planning section alongside the linked project videos. The sequences make intermediate goals easier to understand, while the numerical results show how much harder arbitrary goals remain.
Read the paper