Two new sensing systems, one using ultrasound at the wrist and one using smartwatch sonar, show that the bottleneck in robot dexterity may be sensing, not mechanics.
Two separate research teams published systems that dramatically improve how machines sense human hand movement, using wrist-worn ultrasound and smartwatch sonar respectively.
According to New Atlas, MIT researchers developed an ultrasonic wristband capable of tracking hand movements at a resolution that previous systems could not match. The approach reads muscle and tendon activity directly through the skin using ultrasound. Separately, as reported by Interesting Engineering, researchers from Cornell University and KAIST turned a standard smartwatch into a 3D hand-tracking system by repurposing its built-in speaker and microphone as sonar. Two independent teams, two different hardware approaches, both arriving at the same target: capturing the full complexity of human hand motion in a wearable form factor.
The human hand is mechanically complex enough that no single sensor modality has been sufficient to capture its full range of motion accurately and in real time.
New Atlas puts the challenge plainly: even a routine action involves dozens of small muscles working in coordination. The problem for robotics is that you cannot bolt a force torque sensor onto every tendon. Traditional approaches, optical tracking, gloves with bend sensors, camera-based pose estimation, all make compromises. They lose accuracy under occlusion, they add bulk, or they only capture finger position rather than the underlying muscle state. What the MIT ultrasound approach appears to do differently is read the source signal, the muscle contractions themselves, rather than inferring movement from position data after the fact.
The Cornell and KAIST system takes a different angle. As reported by Interesting Engineering, it repurposes the speaker and microphone already present in a commercial smartwatch to emit and receive sonar signals. The AI layer then interprets the signal reflections to reconstruct 3D hand position. The specs tell a different story about cost here: this approach requires no new hardware at all, just a software and AI layer on top of a device millions of people already wear.
Better sensing is only useful if the actuators receiving the signals can respond with matching precision. These breakthroughs expose how much the sensing side has been lagging behind the mechanical side.
Here is what the data shows: the robotics field has invested heavily in actuator development, torque density, backdrivability, compliance. But dexterous manipulation has remained stubbornly out of reach. The limiting factor may not have been the actuators themselves. It may have been the quality of the input signal those actuators were receiving. If a robot hand does not know what the human hand is actually doing at the tendon level, no amount of actuator refinement closes the gap. The MIT and Cornell work suggests the sensing layer is now catching up.
The Grand Teton robotic sage grouse project shows that deploying actuated robots in uncontrolled real-world environments at small scale introduces a different set of force control and degrees-of-freedom challenges.
According to Interesting Engineering, robotic bird decoys are being deployed at Grand Teton National Park to influence sage grouse mating behavior. The application is unusual, but the engineering constraints are instructive. A robotic bird needs to replicate specific movement signatures convincing enough to trigger behavioral responses in live animals. That requires precise force control and enough degrees of freedom to mimic biological motion at a small form factor, with limited power budgets and no ability to run a power cable into a wildlife habitat. It is a stress test for compact actuator systems under real-world conditions.
All three projects share a common thread: closing the gap between biological motion and machine motion requires better sensing, smarter AI interpretation, and actuators that can respond to richer input signals.
The MIT wristband reads muscle state directly. The Cornell smartwatch system uses AI to reconstruct 3D motion from acoustic reflections. The robotic bird decoy must replicate biological movement convincingly enough to fool live animals. Each project is attacking a different slice of the same core problem: machines still cannot fully observe or replicate the physical nuance of biological motion. What is changing now is the sensing layer. Ultrasound, sonar repurposing, and AI-assisted signal interpretation are all mature enough to be combined into wearable or compact form factors. That is new.
In both the MIT and Cornell systems, raw sensor data alone is not enough. An AI model is required to interpret the ultrasound or sonar signals into usable motion data. This means the sensing breakthrough is actually a sensing-plus-inference breakthrough. The actuator system receiving commands from these sensors is only as good as the model running between the raw signal and the control output.
The next signal to watch is whether any humanoid robot team integrates wrist-based ultrasound or acoustic sensing into their teleoperation or imitation learning pipeline.
Here is what the data shows about the current state: humanoid robot teams are collecting training data through teleoperation, and the quality of that data depends directly on how well they can capture human hand motion. If systems like the MIT wristband or the Cornell WatchHand can be integrated into teleoperation rigs, the quality of dexterous manipulation training data could improve significantly. The robotic bird project is worth watching for a different reason: it will show whether compact actuator systems can sustain convincing biological motion in uncontrolled outdoor environments over time. Both threads feed back into the same actuator question: what does the control input look like, and how precise does the actuator response need to be to match it.
According to New Atlas, the MIT wristband uses ultrasound signals passed through the wrist to read muscle and tendon activity directly. Rather than inferring hand position from finger movement after the fact, it captures the underlying muscle contractions that drive that movement.
As reported by Interesting Engineering, WatchHand is a system developed by Cornell University and KAIST that repurposes the speaker and microphone in a standard smartwatch as sonar. An AI model then interprets the reflected acoustic signals to reconstruct 3D hand position without any additional hardware.
Robot hands can only execute dexterous tasks as well as the input signal they receive. If the sensing system capturing human hand motion is imprecise, the actuators have no accurate target to replicate. Better sensing directly improves the quality of teleoperation data and robot control.
According to Interesting Engineering, robotic sage grouse decoys deployed at Grand Teton require compact actuators with enough degrees of freedom and force control precision to replicate convincing biological motion in uncontrolled outdoor environments. It is a real-world stress test for small-scale actuator systems.
Both the MIT ultrasound system and the Cornell sonar system rely on AI models to interpret raw sensor signals into usable motion data. The sensing breakthrough is inseparable from the inference layer. Raw ultrasound or acoustic data alone does not produce hand tracking without an AI model processing it.