Robotics Briefing /
Robots that adjust, feel contact and remember their footing.
A slight calibration error, a slipping tool or a foothold disappearing from view can undo an otherwise capable robot. Today’s research tests ways to handle those problems. Deployment announcements from a theme park and university campuses show where robots are meeting people, while a recovery benchmark asks what happens after a task goes wrong.
Your quick takeaways
- Self-Adaptive VLA raised average completion under joint-encoder offsets from 5% to 75% across four physical tasks. Each evaluation allowed up to six attempts, with the scene reset between attempts.
- PolyUMI’s combination of vision, touch and contact audio controlled a slipping screwdriver in eight of ten trials, versus two of ten with vision alone.
- AGIBOT announced a rollout of more than 300 robots across Chimelong’s park and hotel operations. The scale is company-reported; sustained uptime and human-assistance rates were not disclosed.
Papers
Self-Adaptive VLA: learning from a miscalibrated attempt
Hongxin Zhang, Chunru Lin and colleagues, UMass Amherst and Genesis AI • Submitted September 24; listed September 25 • Preprint; physical robot tests
A robot can repeat the same miss when its motors move differently from what its software expects. Self-Adaptive VLA learns to use previous attempts as evidence about those hardware errors.
The researchers train with deliberately injected shifts, pair those experiences with corrected expert actions, and add a small encoder that summarizes images, robot state and movements. During deployment, that summary conditions the next attempt; combining summaries helps the policy compensate further.
Across four physical manipulation tasks with joint-encoder offsets, the strongest variant reached 75% average completion, compared with 5% for the base policy. Each task had twenty evaluation episodes, each permitting up to six trials with full scene resets. Under normal calibration, the base policy averaged 88.8%.
This suggests a route to reducing repeated recalibration work. It does not demonstrate continuous repair during a task: adaptation happens between attempts, testing focuses on specific hardware shifts, and the method assumes unsuccessful attempts can be collected safely.
PolyUMI: recording the touch and sound of manipulation
Conor W. Hayes, Rickmer Krohn and colleagues, Northwestern and TU Darmstadt • Submitted September 24; listed September 25 • Preprint; physical robot tests
A camera may miss the moment a tool starts slipping. PolyUMI combines wrist video, optical touch sensing and contact audio in a wireless handheld gripper used to record demonstrations. The same sensing finger transfers to the robot, preserving the sensing arrangement.
In a controlled slip task, a robot loosened and tightened its grasp to rotate a screwdriver from horizontal toward vertical without dropping it. Using eighty demonstrations, the full sensor combination succeeded in eight of ten trials, versus two of ten for vision alone.
The paper also introduces VisTA, a policy that combines information across sensors and time. Its manipulation tests show that additional sensing helps selectively: on lightbulb turning, all methods reached at least 80% success and the vision-only baseline performed best.
The practical contribution is a way to teach and execute skills using contact information that video cannot reliably capture. The release includes collection and deployment software. Small trial counts and a single robot setup limit conclusions about general reliability.
Echo in the Steps: remembering where a humanoid can step
Ming-Ju Lee and colleagues, Tsinghua University • Submitted September 24; listed September 25 • Project reports CoRL 2026; physical robot tests
On a narrow beam or separated stepping stones, a humanoid must place its feet accurately while its camera view keeps changing. This method retains useful depth-image features over time and trains the controller to alternate left and right foot placement.
A Unitree G1 demonstrated traversal over sparse supports, irregular wedges and narrow beams using onboard depth sensing. Reported demonstrations include movement at approximately one meter per second along a twenty-centimeter-wide curved support.
The physical evaluation used ten independent trials per terrain setting. The controller ran onboard at fifty updates per second, showing that the perception and movement pipeline can operate without remote inference.
Remembering terrain could help robots move through constrained spaces where a single camera frame is insufficient. The authors still report trouble with staggered footholds and sharp turns: an apparently valid step can leave the wrong leg supporting the next move, while upcoming footholds can leave the camera’s view. Code is marked as forthcoming on the project page.
News
AGIBOT launches a large robot rollout at Chimelong
AGIBOT • September 24 launch release • Company-reported deployment
AGIBOT and Chimelong announced the first phase of a deployment involving more than 300 robots across an existing theme park and hotel operations. Roles include performances, visitor guidance, educational activities and guest services.
The announcement describes a dedicated 5G-A network with China Mobile and centralized coordination across venues. That makes the project relevant beyond individual robot demonstrations: operating many machines requires connectivity, charging, coordination and support.
For visitors, the intended value is more accessible guidance and interactive experiences. The reported rollout does not establish that every unit operates independently or continuously. The release gives no measured uptime, intervention rate or operating cost.
Robot.com and Grubhub expand campus delivery
Robot.com and Grubhub • September 23 • Catch-up: partnership extension and rollout
The companies announced delivery service for ten additional campuses and a three-year agreement extension. Robot.com says this brings its Grubhub campus presence to more than twenty active locations.
The announcement reports over 25,000 completed deliveries at the University of Southern Indiana, one of the partnership’s original sites. For students, the benefit is a delivery option integrated into the ordering app they already use.
A multiyear extension and a named operating site provide more evidence of repeat demand than a one-off demonstration. The figures remain company-reported and do not disclose delivery cost, completion rates or the amount of remote assistance.
Highlights
RoboRecover: measuring the ability to finish after a mistake
Yang Li and colleagues, Renmin University of China and USTC • Submitted September 24; listed September 25 • Preprint; simulation benchmark
RoboRecover evaluates policies from intermediate states created by earlier mistakes, such as displaced objects or disrupted task progress. It contains two thousand scenarios across RoboTwin and LIBERO, with separate training and test splits.
On LIBERO, UnifoLM scored 98.83% from ordinary starting states but 48.00% on recovery. Another policy, π0.5, scored 94.17% initially and 64.40% on recovery. The stronger starting-state result therefore did not identify the better recovery policy.
This makes recovery an explicit measurement alongside ordinary completion. The study covers two simulators and selected deviation scenarios; it does not establish physical recovery rates or prove which model design caused the differences.
Read this first
Read PolyUMI for a concrete view of what touch and sound add to robot learning. Its slip-control comparison shows a clear benefit in a small physical test, while the lightbulb result shows why adding sensors does not improve every task.
Read the paper