Researchers at the University of California, Irvine have developed TEMPO, a method that helps robot arms handle moving objects by supplying recent visual and action history. Their September 15 paper reports improved bottle handovers, addressing a practical problem: a robot needs to know how a scene is changing to time its grasp. [1]
The project illustrates two gaps in a single camera view. A bottle can occupy the same position while traveling in opposite directions, demanding different interception movements. Likewise, an arm approaching an object can look much like one withdrawing after a completed action. The current image alone may leave both the direction of motion and the stage of the task unclear. [2]
Faster control already addresses part of this problem. The earlier VLASH system lets a robot execute one segment of actions while its model computes another, an arrangement called asynchronous inference. It also accounts for where the robot will be when those new actions start. That scheduling mechanism reduces pauses and aligns predictions with their execution time. [3]
TEMPO adds information about recent events to that control process. Its project description explains how a video encoder, software that extracts visual features, compares the current view with stored features from earlier frames. The resulting motion representation is computed in the background. A separate input summarizes recent joint commands, giving the controller information about its own preceding movements. Both feed a vision-language-action model, which maps camera observations and language instructions to robot actions. Together, these inputs help it interpret motion and distinguish phases of a task. [2]
The researchers tested four tasks on a setup with two I2RT YAM Ultra arms, running 50 trials per method for each task. In BottleHandover, a person walked past while varying speed, bottle height and grip. TEMPO succeeded in 37 trials, compared with 19 for its VLASH baseline. This comparison tests the added temporal information within the study, rather than establishing superiority over the whole field. [1]
The advantage also depended on the task. On DropCatch, TEMPO trailed the authors' implementation of Real-Time Chunking, another method for generating action segments during execution. These laboratory results therefore leave open how reliably the approach transfers to other hardware and operating conditions. [1]
Reproducing that transfer still requires further releases. The official repository contains code, but lists training data, trained checkpoints and deployment support for the tested arms as unfinished items. The present evidence gives robot developers a reason to test visual history and action history together when interception or task sequencing fails. Released models and repeat trials on additional arms would let them judge whether the gains carry over. [4]