Most robot-learning benchmarks begin with a convenient fiction. The workspace is reset, the robot receives one clear instruction and the task ends before much history has accumulated. Real work is rarely so orderly. Instructions change, people move objects, relevant events disappear from view and robots must plan their next action while still completing the previous one.
Three new studies explore what happens when robot learning moves beyond these isolated episodes. ARMS introduces an “always-on” robot policy that watches a continuous stream, remembers its own actions and coordinates two arms simultaneously. OCC4M builds a structured memory of objects and their changing relationships. Streaming-WAM allows a world model to predict and plan while the robot remains in motion.
Together, the research points towards an important shift in physical AI. Future robots will need more than policies that perform individual tasks well. They will also need to preserve relevant history, distinguish important events from irrelevant video and continue operating while their AI systems calculate what should happen next.
From isolated task to continuous stream
Many vision-language-action models are designed around a fixed sequence. A robot receives an instruction, observes the current scene and generates an action. Once the task is complete, the episode ends and the system effectively starts again.
That structure makes experiments easier to design and compare, but it removes several problems that arise during continuous operation. A robot working in a kitchen, warehouse or factory may receive a new instruction while it is still executing an earlier one. An object needed later may be moved out of view. The robot must also remember which actions it has already performed, particularly when two arms are working on different parts of the task.
The CoRL 2026 paper Watch, Recall, Act: Always-On Robots in Concurrent Embodied Streams addresses this with a system called ARMS. It extends the π0.5 vision-language-action model with three lightweight modules that process live perception, the robot’s current physical state and the history of its own actions.
These components update their information asynchronously. The robot does not have to stop moving while it watches the scene or recalls what happened earlier. It also records which arm performed each action and when, allowing the two arms to operate concurrently without losing their individual histories.
In the researchers’ combined evaluation task, ARMS achieved a 45 per cent success rate. The strongest of four main baselines reached 28 per cent. Tests in which individual components were removed indicated that memory, information about the robot’s physical state and asynchronous two-arm execution all contributed to the result.
The absolute performance remains modest. ARMS still failed in more than half of the combined trials, and its “always-on” description refers to the architecture rather than a demonstration of uninterrupted operation over days or weeks. The data were also collected through staged dual-arm teleoperation, meaning that the system has not yet been tested in an unrestricted workplace.
Even so, the research captures a problem that conventional benchmarks largely avoid. A continuously operating robot cannot assume that each new instruction arrives in a clean environment. It must interpret the present in relation to a history that it has partly created itself.
Why more video does not necessarily produce better memory
Giving a robot access to its complete video history might appear to solve the memory problem. In practice, a growing archive of frames is expensive to process and does not clearly distinguish significant events from visual repetition.
A robot may need to know that a container was moved, which object was placed inside it or whether two identical-looking items are actually the same object seen at different times. Recovering that information from hundreds of raw images is not the same as maintaining an explicit record of the world.
OCC4M, pronounced “Occam”, takes the latter approach. The preprint describes an object-centric four-dimensional memory that tracks objects in a common three-dimensional coordinate system over time. In addition to locations and identities, the memory represents relationships such as movement and containment.
A vision-language model can query this structured representation when it needs to select an object or action. The low-level manipulation policy does not need to receive the full visual history.
Across 350 simulated episodes, OCC4M achieved a memory success rate of 96.6 per cent and an end-to-end task success rate of 88.9 per cent. A baseline that supplied Gemini 3.7 Flash with the full observation history reached 54.6 per cent memory success and 57.7 per cent end-to-end success.
The difference became particularly pronounced when the viewpoint changed. OCC4M retained 100 per cent memory success and completed 98 per cent of the tasks. The full-history approach fell to almost zero. According to the researchers, representing objects in a shared world frame allowed the system to preserve their identity even when the camera saw the scene from a different position.
The physical results were less conclusive. In twenty trials with a Franka robot, OCC4M achieved 85 per cent accuracy on the combined memory questions, compared with no more than 30 per cent for the frame-based approach. Yet the robot completed only 45 per cent of the full two-stage tasks.
This gap is important. Remembering the correct object or relationship does not guarantee that the robot can successfully execute the necessary movements. The physical evaluation was also small and used a fixed camera. The study provides stronger evidence for structured robot memory than for dependable long-horizon manipulation.
OCC4M nevertheless reinforces a finding that is appearing across several recent studies: storing more visual history is not automatically the best way to remember. A compact and explicitly structured representation can be both more efficient and more robust than repeatedly asking a large model to search through raw footage.
Thinking while the robot keeps moving
Memory is only one part of continuous operation. A robot must also avoid long pauses while its models process observations and calculate future actions.
This is becoming particularly relevant for world-action models. These systems predict how the scene will change in response to an action and use that forecast to determine the robot’s next movement. The additional reasoning can improve control, but generating future visual states is computationally expensive.
In a synchronous system, the robot completes an action segment and waits while the model makes the next prediction. The result can be a pattern of moving and stopping that slows down even relatively simple tasks.
Streaming-WAM attempts to remove this delay by overlapping calculation and execution. While the robot performs an already committed part of an action sequence, the model predicts the future scene and generates the remaining actions. Crucially, the prediction includes the movements that have already been scheduled but not yet completed.
Without this action conditioning, the model would plan from an outdated view of the world. By the time its calculation finished, the robot and the objects could already be in different positions.
On the LIBERO manipulation benchmark, Streaming-WAM achieved an average success rate of 98.35 per cent. It reduced mean episode duration by a factor of 2.93 compared with Fast-WAM. In a physical task in which a robot placed a stamp on paper, average execution time fell from 90 seconds with a synchronous world-action model to 38 seconds.
These results show the potential importance of asynchronous control, but the physical evidence remains narrow. A single stamping task does not establish that the same acceleration is safe during complex assembly or unpredictable human-robot interaction. The approach also commits the robot to executing part of an action sequence while the next prediction is still being calculated. Those committed actions must therefore remain safe even when the environment changes unexpectedly.
The broader principle is nevertheless significant. In physical AI, inference time is not merely a computing benchmark. It directly affects how a robot behaves. A capable model that repeatedly forces the machine to pause may be less useful than a slightly simpler system that can continuously observe, plan and act.
Continuous operation changes the research problem
ARMS, OCC4M and Streaming-WAM address different parts of the same challenge. ARMS treats robot operation as a continuing stream rather than a series of independent instructions. OCC4M converts visual history into a persistent model of objects and events. Streaming-WAM overlaps prediction with movement so that deliberation does not repeatedly halt the task.
None of the three systems is ready for unsupervised, continuous industrial deployment. ARMS has a success rate of 45 per cent on its combined task. OCC4M remembers more reliably than it manipulates. Streaming-WAM has limited real-world validation and must manage the safety implications of committed actions.
Their importance lies in the research direction they share. As robot policies move into longer and more realistic deployments, success will depend on more than completing an action from a predefined starting state. Robots will need to decide what to remember, relate new observations to their own earlier behaviour and continue working while their models reason about what comes next.
That shift also changes how systems should be evaluated. A robot that succeeds after every scene has been reset may still fail when objects move between instructions or when its own previous actions alter the workspace. Likewise, a model that performs well in terms of task success may be too slow for practical use if it introduces long pauses between movements.
The next generation of physical AI will therefore have to be judged not only by what robots can do, but also by whether they can maintain context and momentum. The goal is no longer simply a robot that performs a task. It is a robot that keeps track of an unfolding world—and keeps working while it thinks.
Sources
Watch, Recall, Act: Always-On Robots in Concurrent Embodied Streams
OCC4M: Object-Centric 4D Memory for Spatiotemporal Reasoning in Long-Horizon Manipulation
Streaming-WAM: Action-Conditioned World-Action Model for Asynchronous Robot Manipulation
