UMArm
UMArm · pneumatic soft robot arm

Keep the compliance.Fix the performance.

Soft robots
LightCompliantFastPreciseStrong
UMArm aims for
LightCompliantFastPreciseStrong

The bet: many actuators, each precisely controlled, packed densely into one arm.

valves + controller in the cap pressure, tracking its setpoint illustration
01

Valves go inside the muscle

IROS 2024
Top: an assembled valve-embedded McKibben actuator. Bottom: the parts that sit inside it, two valves joined by a flexible circuit board. outside inside valve pressure sensor + controller on a flex PCB valve

Before. One tube per muscle, regulator far away.

Now. One air line and four wires for all of them.

Each muscle carries its own valves, pressure sensor, and controller. Any number of them share one air line and four wires.

  • DidTested against an external regulator. No visible delay, where a 100 cm tube adds 100 ms.
  • LimitSmall commercial valves cap how fast pressure can change.
37–53%
lower pressure-tracking error
vs. an external regulator through 100–25 cm of tube
±0.05psi
pressure deadband
over a 0–40 psi working range
400Hz
pressure loop, in every muscle
valves switch in under 5 ms
02

UMArm: 24 muscles on one spine

arXiv 2025
The first UMArm hanging from a frame, slightly bent, its grey and black muscles wrapped around a jointed spine
Wearable third arm
Untethered inchworm
Directional stiffness
Stacks within 3 mm

A 1.15 kg arm: a jointed spine pulled by 24 self-regulating muscles. Opposing pairs set where each joint points and how stiff it is.

  • DidForce, accuracy, and stiffness tests. Four demos, including an untethered inchworm.
  • LimitStiffness was set only at the straight pose. Off-the-shelf joints limit the load.
1.15kg
with all 24 regulators on board
comparable soft arms: 1.0–1.6 kg, tethered
53N
peak pull at the tip
comparable soft arms carry 0.3–0.5 kg
0.85mm
steady-state tip error
30 mm for a Koopman-controlled soft arm
9.6×
stiffer along one axis
37.5 vs. 3.9 N/m, measured at one pose
03

ProMax: fast and compliant together

arXiv 2026 · experiments ongoing
targetpen, motion capture
0.75m/s
pen speed
stiffness ellipse steered along the path
21.2mm
RMS error, 0.72 m wide word
earlier controller: 27–31 mm

Writes a word in the air at up to 0.75 m/s, while a learned model predicts and steers how stiff the tip is.

commandpressuretensionmotion + stiffness
  • DidCustom joints, 24 muscles on a 150 Hz bus. A three-stage model, with sampling MPC replanning at 50 Hz.
  • LimitStiffness in motion is a model estimate, not yet probed by force. Gravity still dominates it.
6.1mm
RMS error at 0.19 m/s
earlier Koopman-model controller: 9.4 mm
4–5mm
tip prediction, 0.4 s ahead
a hold-still guess: 22–33 mm
3.6ms
for 1024 model rollouts
inside a 20 ms planning period

The boards: one controller per muscle

VEMA board family
Thirteen circuit boards laid out in a row on a grey mat: a ring-shaped board still in its panel, a long green board, an orange flex board, nine small rectangular boards, and a large ring board with white and green connectors flex board V6 Rev B · 35 × 12 mm The family so far · photo

VEMA, the valve-embedded McKibben actuator, needs a board in every muscle: valve drivers, a pressure sensor, a processor, and a bus.

  • DidFlex and rigid boards, one revision after another. V6, the latest, drives four valves and talks CAN FD.
  • LimitV6 has run its pressure loop on one actuator so far.
35 × 12mm
V6 Rev B: 67 parts on 4 layers
four valve outputs; the 2024 flex board drove two
5Mbit/s
CAN FD data rate on V6
the CAN boards before it: 1 Mbit/s
500Hz
bus cycle V6 was tuned at
the ProMax arm's bus runs at 150 Hz

All demos and papers