Put the big motor on the finger that has to hold. The Berkeley QUAD hand gives dexterity to small quasi-direct-drive motors and sustained grip to one larger motor, and the paper reports up to a 96-fold cut in heat during sustained loading. Agent systems should size for duty cycle the same way: small and sensitive for the short work, bigger and cooler for the hold.
There’s a photo in my head of a robot hand on a lab bench with a tiny bandage on one finger. Somebody pushed it too hard holding something too long.
That’s a duty cycle problem. The Berkeley QUAD hand is a clean answer to it.
What is the Berkeley QUAD hand?
The QUAD hand paper from Benjamin Davis, Chase Kidder and Hannah S. Stuart came through Andra Keay’s Robots & Startups reading list from the Bay Area Robotics Symposium. QUAD is Quasi-direct-drive, Underactuated, Asymmetric Design. Four fingers, 11 degrees of freedom, 8 degrees of actuation, about 1 kg.
They dropped the fifth finger because its motion correlates with the ring finger in grasping. They also dropped distal IP joints, a common actuator-saving choice. The index finger keeps MCP abduction so it can move relative to the thumb and middle. The underactuated ring finger skips abduction and gets a CMC joint for cupping and closure.
Quasi-direct-drive motors give great force feedback. You can feel contact through them, and the paper reports backdrive forces as low as 50 g on the thumb and index base joints. The catch is low torque density, which means size, weight and heat. Holding a grip with a small QDD motor means pushing current through it continuously. Fully QDD five-finger hands run into packaging and thermal limits, which is why earlier QDD designs the paper cites stay at three fingers.
So they stopped treating every finger the same:
- Thumb and index each get their own small SteadyWin GIM3505-08 motor. These are the dexterity fingers.
- Middle and ring share one larger CubeMars AKE60-8 motor through a preloaded, compliant linkage. Under light load the middle finger leads and the ring finger mirrors it. Under heavy load the middle finger complies and the motor drives the ring finger down and inward for better closure.
The paper gives a force threshold of about 3.2 N at the middle fingertip for that switch, tunable by spring preload.
Sharing one motor across two fingers frees up space in the palm, and that space is what lets a bigger motor fit. Distal joints on the first three fingers still use high-reduction Dynamixel servos, with a KST A08 micro servo on index abduction. QDD stays near the palm, where backdrivability covers the widest range of use, and distal mass stays down.
Why does putting the big motor on two fingers cut heat by up to 96 times?
I checked this one against the paper before writing it. The abstract says the design yields “up to a 96-fold reduction in heat generation during sustained loading.” The body explains why.
Table I lists the motors. The GIM3505-08 is 43 mm across, 97 g, 8:1, 0.65 Nm rated, phase resistance 5.06 ohms. The AKE60-8 is 69 mm across, 260 g, 8:1, 5 Nm rated, phase resistance 0.289 ohms. Torque density goes from 6.70 Nm/kg to 19.2 Nm/kg. The small motor’s phase resistance is more than 17 times the larger one’s, because a smaller radius needs more turns of thinner wire.
They then held a 3 N fingertip load for five minutes, and a 6 N load with both ulnar fingers. Heat is calculated from quadrature-axis current and that phase resistance, not from a thermocouple on the case. Combined with lower current draw, that gets to “up to 96×” with the shared actuator, “even when normalizing for force per finger.”
The ulnar motor drew the least current. It drew about twice as much through the middle finger as through the ring finger, which they attribute to finger length and linkage kinematics. Loading both ulnar fingers at 6 N drew current similar to loading the middle finger alone at 3 N, which suggests most of the load transferred through the ring.
Note the “up to.” It’s a best case under sustained loading, not an average across every grasp.
Backdrive is the other side of the split. Thumb and index sit around 50 g. Middle and ring sit near 250 g, still usable for a lot of everyday grasping, not the same as the dexterity pair. That is the trade they accepted on purpose.
The hand still covers 29 of 33 Feix taxonomy grasps, so it didn’t trade away much range to get there. Missed grasps were Light Tool (no DIP actuation), Sphere 4-Finger and Prismatic 4-Finger (no fifth finger), and Prismatic 3-Finger (no independent ring). It also hits 10 of 11 Kapandji opposition postures, missing the one that needs a fifth digit.
They are careful about what they did not measure. The taxonomy and Kapandji tests are static poses, not dynamic manipulation. Force tests used one position per joint. They did not compare peak fingertip force, because the 3D-printed ulnar linkages yield before the big motor stalls. A supplemental video shows lifting a bulky object of about 5 kg as an example of offloading onto the ulnar side, not as a rated spec.
The heat number is a calculated resistive-loss ratio on a five-minute hold, with the motors they named. It is not a sticker that says 96 times more efficient at everything.
Why should agent systems size for duty cycle the same way?
I spend most of my time on agents rather than hands, and this maps over almost directly.
Most agent stacks size every component the same way. Same model, same context budget, same polling interval, for the quick reactive work and the long sustained work alike. That’s five identical small motors.
QUAD’s split is the better pattern:
- Dexterity work is short, varied and needs feedback. Routing a message, picking a tool, answering a question. Small, fast, sensitive components are right here.
- Holding work is long and steady. Watching a queue for hours, keeping a long-running session, re-reading the same reference every wake-up like my house agent and its rulebook. Running that on the dexterity components burns money the way a small motor burns current holding a grip.
Eight hours, not one task is the eval version of this. A one-task demo is a precision pinch. A shift is a hold. The 20-hour robot is the same split on batteries. Run the model on the quiet chip is the compute version: put the always-on part on the silicon that can sit there without cooking.
So size for duty cycle. Put the sustained load on the component built for sustained load, even if it’s less nimble, and let the small ones do the fine work.
Then you don’t need the bandage.
What duty-cycle checklist would I use on an agent stack?
If you are about to run every step on the same model, walk this before you ship the bill.
- Label each step dexterity or hold. Dexterity: one-shot routing, tool choice, a short answer. Hold: queue watchers, session memory, rulebook rereads, overnight soaks. If you cannot label it, you will default to the expensive motor.
- Write the duty cycle. Wakes per hour, minutes of continuous context, how long a grip has to last. The house agent that rereads a long prefix every wake is a hold, even if each wake looks small.
- Put holds on the bigger, cheaper path. Smaller model, cached prefix, less polling, a worker that is allowed to be slow. QUAD put torque and heat capacity on the ulnar side because that side closes and stays closed.
- Keep a sensitive path for contact. The 50 g backdrive is the point of QDD on thumb and index. Agents still need a fast, high-feedback path for “something just changed.” Do not collapse that into the hold path.
- Do not average the heat number. The paper’s 96-fold cut is sustained loading, five minutes, named motors, “up to.” Report the hold case separately from the demo case.
- Measure current, not just torque. They computed heat from quadrature-axis current and phase resistance. For agents, that is tokens per hour while idle-but-present. Eight-hour evals catch this. One-task evals do not.
- Watch the part that yields first. Their 3D-printed linkage is the limiter, not the big motor. In an agent stack the limiter is often the rulebook length, the session store, or the retry loop. Fix the part that is actually hot.
A symmetric hand of small motors looks elegant on a slide. It runs hot in a clenched grip. Symmetric agent stacks do the same thing with tokens.
Related
- Source: A Quasi-Direct-Drive Underactuated Asymmetric Hand for Dexterous and Efficient Grasping and Manipulation, Davis, Kidder, and Stuart, arXiv:2609.27240
- Project page: Berkeley QUAD Hand
- Also in: Robots from the customer down, not the task up, Andra Keay
- The hard part is the hands
- Your house agent pays to reread its rulebook
- Eight hours, not one task
- The 20-hour robot
- Run the model on the quiet chip