Robots need better timing—and a way to know when things go wrong.

Chronos Policy: the right action needs the right time

Changhao Wang, Jaden Clark and Shuran Song, Stanford • Submitted October 9; listed October 12 • Preprint

A camera frame tells a robot where an object was when the image was captured. If sensing or computation slows down, acting as though that image is current can turn a good grasp into a miss. Chronos Policy makes both observation times and intended action times explicit inputs to a learned controller.

During training, the researchers vary sensor sampling rates, delivery delays and when a predicted sequence of actions begins. The model can then distinguish observations separated by different amounts of real time, even when they occupy the same positions in its input sequence.

In physical tests, a robot placed a cup onto a moving saucer, inserted a peg and pivoted objects upright. For the moving cup under normal timing, Chronos achieved eight full successes in ten trials, compared with five for Diffusion Policy. With sensing and inference delays combined, it achieved five full successes and one partial success; the baseline achieved three full successes.

Those are small experiments, with ten trials per cup condition. The paper’s percentage scores also award half credit for partial success, so they are not always completion rates. The broader implication is that deployment reliability depends on knowing how old information is. Variable control frequencies remain future work, and the project’s code is still marked coming soon.

• Paper ↗
• Project and experiments ↗

Robot-WATCHDOG: spotting a task that looks normal but is going wrong

Quentin Rolland and colleagues, CEA List, Inria and Bleu Robotics • Submitted October 9; listed October 12 • Accepted at CoRL 2026

A robot can follow a familiar movement while putting a cup in the wrong place relative to a plate. Robot-WATCHDOG represents a scene as tracked objects and the changing relationships between them. It compares those relationships with successful demonstrations to flag mistakes.

The study evaluates six manipulation tasks across two datasets, including table setting and conveyor-belt waste sorting. Its simpler scoring method, TPC, outperformed a learned graph predictor and the tested alternatives on the main dataset averages. After converting scores into alerts, its balanced detection measure reached 0.675 on BotFails, versus 0.220 for the strongest online baseline there. This is a detection metric, not a robot task-success percentage.

The useful distinction is between movement that merely looks unusual and movement that breaks a task’s structure. A separate monitor could help identify failures that a controller itself does not recognize. These results evaluate detection on recorded robot episodes; they do not establish that interventions prevented damage or recovered failed tasks.

The representation tracks object centers, so it misses orientation errors and does not naturally cover liquids or highly deformable objects. Occlusion and tracking mistakes can also corrupt its alerts. The project provides qualitative examples; a complete downloadable code or dataset release was not verified.

• Paper ↗
• Project and examples ↗

SAS: helping a tracked robot move when its map is misleading

Valentýn Číhala and colleagues, Czech Technical University in Prague • Submitted October 9; listed October 12 • Preprint

Tracked robots use articulated extensions called flippers to climb obstacles and keep their bodies stable. Many controllers rely on an elevation map, including ground that the robot’s own chassis now hides from its sensors. Position-estimation errors can gradually make that remembered terrain inaccurate.

Stabilizing Anti-Stuck, or SAS, instead uses body tilt, flipper angles and a short-term indication of whether the robot is moving. It tries to keep the body level and changes flipper motion when progress stops.

In a physical demonstration, SAS crossed a log concealed by tall grass. A map-based comparison mistook the visible grass geometry for a drop, lowered its flippers and became stuck. That is a concrete example of simple feedback remaining useful when perception is wrong.

The measured pallet evaluation was more mixed: SAS ranked second among autonomous controllers using the onboard-map condition, while manual control scored higher than every controller. Each condition comprised one ride of ten crossings, and some baselines received SAS-based anti-stuck assistance. Without terrain preview, SAS also reacts late to obstacles beneath the chassis. The result supports a useful fallback strategy, with limited evidence for broader field reliability.

• Paper ↗

An Arctic robot fleet maps a glacier above and below the water

October 9 • Catch-up • British Antarctic Survey report on summer fieldwork

British Antarctic Survey reports results from a thirty-day expedition to Greenland’s Sorgenfri glacier. Surface and underwater robots produced thirty glacier-face maps, paired with drone measurements above the water. The team also sent Boaty McBoatface beneath kilometers of floating glacial debris.

Different machines supplied different views: DriX surveyed near the surface, Gavia and ecoSUB worked underwater, and a remotely piloted boat carried a tethered vehicle that placed an instrument into the ice. That instrument returned nineteen hours of measurements close to the glacier face.

The value is access to places and measurements that are difficult or dangerous for people to reach. This was a coordinated scientific operation involving autonomous vehicles, remote control, a research ship and skilled staff. Researchers expect the data to improve glacier models over the next three years; the report does not yet demonstrate more accurate sea-level forecasts.

• Expedition report ↗