Researchers tested ARMS, a dual-arm robot controller that records what each arm has done so it can respond to later instructions while work continues. In the paper's combined-task evaluation, it reached 45% end-to-end action success, compared with 28% for SwitchVLA, the strongest of four baselines.
ARMS is designed for an ongoing stream rather than a sequence of commands separated by resets. A standing rule can trigger later; a new request can arrive while the current job continues; and an instruction can refer to an object the robot moved earlier. With both arms sharing one scene, the controller must track their actions and avoid conflicting paths.
ARMS keeps a pretrained vision-language-action model as its action engine and feeds it context from three lightweight modules. A visual trigger watches the camera stream and wakes the main controller when a standing rule is met. A state module reports what each arm is doing and whether their planned paths may collide. A compact memory records the arm, object and time associated with each action.
These modules update asynchronously, so the robot can keep moving while they watch and retrieve earlier actions. That record can answer a later request such as "put the second one you stowed in the box" without relying on the current image alone. It can also distinguish two similar objects by the order and arm that moved them.
The 45% result is the study's combined-task action-success score, not a general robot success rate. An episode counted only if the foreground task continued, every triggered standing rule was handled by a free arm without collision, memory retrieved the correct earlier object, and the prescribed priority order was followed. The four baselines shared the same pretrained backbone, fine-tuning data and backbone-training budget; ARMS also trained its added modules.
In a separate live-robot test on a Cobot Magic dual-arm platform, ARMS's long-stream retention score was 67% for streams lasting at least 10 minutes without a reset, compared with 30% for the strongest baseline. Across three randomized rounds, the authors report 60 trials per method for each of four segments: watch plus concurrency, recall, self-perturbation, and long streams.
The researchers built the ARMS Dataset from about 200 staged, never-reset streams. The full-overload test, in which both arms are busy when a new standing rule triggers, was covered mainly in one task family. Evidence remains small-scale and limited to one platform, with standing rules predefined and priority tiers fixed in advance. The authors call for broader real-robot data, learned standing rules and priorities, and richer overload policies.
Sources
- Original paper: "Watch, Recall, Act: Always-On Robots in Concurrent Embodied Streams"
- Author publication page listing the paper as accepted to CoRL 2026
- Lead image: Figure 3 from the original paper, used under CC BY 4.0.