Robotics Briefing /
Learning to recover, sharing what robots see, and moving around trouble.
Robots encounter situations their demonstrations did not prepare them for: a folded cloth goes wrong, a pedestrian appears from behind a wall, or a narrow passage demands a change in posture. Three new papers examine ways to handle those moments. Beyond the lab, forestry automation is expanding and a rover has begun its journey toward Phobos.
Your quick takeaways
- TILDE learns from a stronger robot policy’s corrections. It recovered from eight of twelve predefined cloth-folding failures, versus four for the demonstration-only baseline.
- CoopUQ shares observations between robots to anticipate hidden pedestrians. Physical navigation success averaged 94% across 48 controlled trials, with external localization and computing.
- SWAP lets a humanoid duck, sidestep and adjust its arms around obstacles. Physical tests succeeded in three or four of five attempts per scenario; dynamic obstacles were tracked with motion capture.
Papers
TILDE: the recovery can be as useful as the demonstration
Yuxiang Lu, Yantong Liu and colleagues, University of Hong Kong • October 18 preprint
Learning from successful demonstrations leaves a robot poorly prepared for the mistakes it makes while acting. TILDE adds experience from those mistakes and from a stronger model’s attempts to correct them.
A progress monitor watches the student policy. When progress stalls, a teacher policy takes control and attempts a recovery. Successful recoveries and the preceding student actions become training data, alongside human demonstrations and teacher guidance. The student avoids repeating large-scale robot-action pretraining, although it still starts with a pretrained vision-language backbone.
On a physical cloth-folding test with twelve predefined failure states, TILDE recovered successfully in eight, compared with four for behavior cloning and nine for the teacher. Across cloth folding, pen pickup and beverage packing, it also improved scores that award partial task progress. The headline 4.1-times improvement refers to those scores, not a fourfold increase in complete-task success.
The useful finding is that executing a correction supplies experience that merely labeling the student’s next action misses. The evaluation remains small and task-specific, and the approach depends on both an accurate progress monitor and a teacher capable of fixing the failure.
CoopUQ: another robot may see the pedestrian hidden around the corner
Zehao Wang and colleagues, Georgia Tech and collaborators • October 19 arXiv release • CoRL 2026
A robot cannot reliably predict a person it cannot see. CoopUQ combines information from neighboring robots to forecast which nearby spaces will be occupied, including areas hidden from an individual robot’s view.
Each robot summarizes its recent laser scans and shares that representation. The navigation system considers both predicted occupancy and uncertainty, penalizing routes through areas likely to be blocked or poorly understood. This helps it slow down or detour before a pedestrian fully emerges.
Two physical ROSMASTER X3 robots were tested in four indoor environments with one to three pedestrians. Across 48 trials, the reported average success rate was 94%, versus 58% for SCOPE. With three pedestrians, success fell to 88%, and collisions still occurred. The policy transferred from simulation without additional physical training.
Sharing observations offers a way to improve visibility without putting every sensor on one robot. These tests used external ultra-wideband localization and a laptop GPU, however, and involved matching robots with 2D laser scanners. They do not establish performance for mixed fleets, dense public crowds or unreliable communications.
SWAP: collision avoidance involves more than choosing footsteps
Zeji Yi and colleagues, Carnegie Mellon University and Amazon FAR • October 19 preprint
A humanoid can fit through a passage by lowering its body, turning sideways or moving an arm. SWAP separates choosing such movements from the controller responsible for maintaining balance.
A generative planner proposes possible motions, while a learned controller tracks them. Training examples are created by placing simulated obstacles outside the space a recorded motion occupied. At deployment, candidate motions can be scored for preferences such as posture or direction without retraining the static-obstacle policy.
In physical Unitree G1 trials, success was four of five for sidestepping, three of five for ducking, four of five for lifting a hand, and three and four of five for fast and slow moving obstacles respectively. These small hardware tests are more informative about current readiness than the separate 88.3% average simulation success rate.
The physical setup used laser sensing for static obstacles, but motion capture supplied robot localization and moving-obstacle states. Dynamic tasks also used a separately trained planner. The results show several useful avoidance behaviors; they leave open how reliably the system works in unfamiliar environments without that external sensing support.
News
Kodama raises $7 million as forestry machines work under remote supervision
October 19 announcement • Company-reported operations
Kodama announced $7 million in financing, bringing total funding to $14 million. Its systems combine supervised autonomy for repeatable work with remote operators handling complex situations, while integrating with existing forestry equipment.
The company says equipped machines moved more than 100,000 tons of timber across southeastern U.S. logging operations over six months. That is evidence of reported commercial activity, but it does not measure how much work was autonomous. The announcement gives no intervention rate or independently measured productivity comparison.
The funding will support more deployments, equipment partnerships and autonomy development. The operational question is whether remote supervision can keep expensive machines productive while making scarce operator time go further.
Idefix begins its journey toward a very different driving environment
October 19 launch • DLR mission update
Japan’s MMX spacecraft launched with the German-French Idefix rover aboard. DLR says the rover is scheduled to reach Phobos’s surface around the turn of 2028/2029 and explore for about 100 days.
After release above the surface, the plan calls for it to bounce, settle, right itself autonomously and deploy its solar panels. Gravity below one-thousandth of Earth’s makes even driving and maintaining contact unfamiliar control problems; long communications delays also limit direct intervention.
The launch is the completed milestone. Landing, self-righting and surface exploration remain future tests. Successful operation would provide both robotics experience in extreme low gravity and surface measurements to inform the mothercraft’s later sampling.
Highlights
MoveIt Pro adds tools for checking a learned policy before deployment
October 20 release notes • Commercial software
MoveIt Pro 10.2.0 documents a repeatable evaluation workflow: reset a task, run the policy, score the outcome and report a success estimate with uncertainty. It also adds episode-by-episode review of imported LeRobot datasets. Those features make it easier to inspect the evidence behind a trained robot’s performance. The quality of the evaluation still depends on the chosen resets and scoring rules; a high score on one setup cannot establish reliability elsewhere.
Read this first
Watch CoopUQ’s comparison of robots navigating with and without shared observations. The hidden-pedestrian example makes the benefit concrete: another robot’s view can provide time to react before a person emerges. The physical test results and remaining collisions provide useful context alongside the videos.
Read the paper