The rapid growth of physical AI is often presented as a race to build larger models and train them with more data. Three new studies suggest that this is only part of the equation. Robots also need hardware that can respond quickly to physical contact, memory systems that do not slow down every movement and sensors that reveal what is happening when cameras no longer provide a clear view.
Research into the AthenaZero robot, Workspace Models and the tactile Agile-WAM system approaches the problem from three different directions. Together, the studies illustrate a broader shift in robotics research. Better AI remains important, but progress in the physical world increasingly depends on how models interact with robot bodies, limited computing resources and uncertain contact.
Rethinking the robot arm
AthenaZero starts with the mechanics. The two-armed research robot was designed to minimise the effective mass of its arms while retaining enough power for fast and dynamic manipulation. Its developers used quasi-direct-drive actuators, relatively low gear ratios and transmissions that place much of the motor mass closer to the centre of the robot.
That differs from many conventional industrial and collaborative robots. These generally use highly geared actuators to provide stiffness, precision and the ability to hold a load in one position for long periods. The disadvantage is that the resulting arms have considerable reflected inertia. They can be accurate and powerful, but their physical response to an unexpected impact is very different from that of a human arm.
AthenaZero was tested through a series of baseball-inspired demonstrations. It threw balls at speeds exceeding 30 metres per second and caught and struck balls travelling above 14 metres per second. These tasks require the robot to accelerate rapidly and react within milliseconds.
The more significant evidence comes from a controlled impact test. At the wrist, AthenaZero had a measured effective mass of 0.83 kilograms. A Franka Research 3 arm measured 3.3 kilograms under the same test conditions. When struck by a pendulum, AthenaZero also produced a lower peak impact force: 124.1 newtons, compared with 208.3 newtons for the Franka.
These results support the researchers’ argument that dynamic manipulation is partly a hardware problem. A controller cannot fully compensate for the physical inertia of an arm at the initial moment of impact. Making the mechanism itself lighter and more responsive can therefore provide capabilities that cannot simply be obtained by installing a larger AI model on an existing robot.
The design involves clear trade-offs. AthenaZero has less passive stiffness and lower continuous holding torque than conventional industrial arms. It is mounted on a fixed base, draws power from an external supply and uses hands developed partly around the baseball experiments. The system is not a direct replacement for the robots used in production cells.
Its importance lies instead in the design question it raises. If robots are expected to catch objects, handle unpredictable contact or work safely at human speeds, hardware optimised primarily for rigidity and static precision may no longer be sufficient.
Memory without a large model in the control loop
The Workspace Models study addresses a different physical constraint: computing time. Robots performing longer tasks need to remember previous events. A robot may have to recall which drawer contains an object, how many items it has already moved or what happened during an earlier attempt.
One approach is to give a vision-language model access to the robot’s image history and ask it to select or summarise relevant information. This can provide useful context, but querying a large model during execution introduces latency and requires substantial computing capacity. For a robot that must react continuously, even a relatively short delay can disrupt control.
Workspace Models moves this expensive reasoning step to the training process. During training, vision-language models identify important events and visual details. That information is then distilled into a compact internal representation called a workspace token. When the robot is deployed, its policy can consult this learned state without repeatedly running the original vision-language model.
The method was evaluated on three simulated tasks and one task using a physical Franka robot. The experiments tested different forms of memory, including counting objects that disappeared from view, recalling a previously opened drawer and learning from an earlier attempt to balance an object.
Across the four tasks, Workspace Models achieved an average success rate of 91.5 per cent. The strongest keyframe-based comparison reached 66.8 per cent. In its most optimised form, the workspace system required 49.2 milliseconds for inference. The comparable batched keyframe method took 199.5 milliseconds, while a policy without equivalent memory took 41.3 milliseconds.
The results indicate that compact memory does more than reduce computing costs. Feeding a policy a long sequence of historical images can introduce irrelevant information and interfere with precise control. A smaller representation of the important events may consequently work better than a richer but less selective history.
The scope of the evidence remains limited. The researchers designed four specific tasks, and the training process still depends on vision-language models and predefined questions about what the robot should remember. It is not yet clear whether a similar architecture could maintain reliable memory throughout an open-ended factory or warehouse process.
Nevertheless, the study points towards a practical division of labour. Large models can help determine what matters during training, while a smaller system carries that knowledge into deployment. That could make embodied AI less dependent on expensive onboard hardware and external model services.
Feeling when an insertion goes wrong
Agile-WAM focuses on another weakness of visually controlled robots: cameras often become least useful when physical contact becomes most important. During insertion or assembly, the gripper and component can obscure the target. A camera may show that the robot has reached the correct area without revealing whether a peg is aligned, a connector has locked or two gear teeth are engaged.
The system combines images with tactile measurements and trains a world-action model to predict both future sensor states and future actions. The objective is not merely to recognise contact, but to learn how contact is likely to develop and which corrective movement should follow.
Researchers evaluated Agile-WAM on nine simulated tasks and five physical tasks involving gear assembly, peg insertion, power plugs and Ethernet connectors. During the hardware experiments, the robot sometimes began from a deliberately imperfect alignment, requiring the policy to recover rather than simply execute a successful trajectory from start to finish.
Agile-WAM completed gear assembly in 16 of 20 attempts. The strongest visual-tactile comparison succeeded 11 times. For power-plug insertion, it achieved 13 successes, compared with seven for that comparison. During Ethernet-cable removal, it succeeded in 17 of 20 trials.
The model could use contact information to slide a misaligned peg towards a hole or rotate a gear until its teeth engaged. These are small corrections, but they address a central problem in industrial automation. A robot following a fixed trajectory may stop or cause damage when reality deviates slightly from its programmed assumptions.
The results were not uniformly strong. Agile-WAM performed slightly worse than the best comparison on peg insertion, while Ethernet insertion succeeded in only six of twenty attempts. Each physical task was also evaluated over just twenty trials and trained with fifty task-specific teleoperated demonstrations. Its low inference time of 10.35 milliseconds was measured on an Nvidia RTX 4090, so efficient deployment on smaller industrial edge computers has not yet been demonstrated.
Even with these qualifications, the work shows why tactile sensing is becoming increasingly relevant to physical AI. Touch is no longer treated solely as an additional safety signal. It is becoming an input through which learned policies can understand contact, predict what is going wrong and select a recovery action.
The physical system matters
AthenaZero, Workspace Models and Agile-WAM solve different problems, but they share the same underlying lesson. Performance in robotics cannot be separated from the physical system in which an AI model operates.
A robot with heavy, highly geared arms cannot respond to contact like a low-inertia mechanism, regardless of the sophistication of its software. A policy that repeatedly consults a large model may possess more contextual information but react too slowly. A camera-based system may recognise the correct object and position yet remain unable to determine whether two components are properly connected.
The next stage of physical AI is therefore unlikely to be driven by model scale alone. It will depend on the co-development of hardware, sensing, memory and control. Larger models can still provide general knowledge and broad capabilities, but robots need mechanisms that can act on that knowledge within the time and contact constraints of the physical world.
For industrial automation, that could lead to robots that are less rigid in both the mechanical and operational sense. They may react more safely to impacts, remember earlier steps without expensive online reasoning and recover when assembly does not proceed exactly as expected. The three studies do not yet deliver that complete system, but they make increasingly clear what its components will need to look like.
Sources
- AthenaZero: A Low-Inertia, Bimanual Robot for Dynamic Manipulation
- Workspace Models: Lightweight Robotic Memory via Saliency-Driven Supervision
- Agile-WAM: An Agile Tactile World Action Model for Contact-Rich Robot Control
