Unscrewing a tube cap by feel

Before anything can be poured, the cap has to come off. One UR5e holds the tube steady while the other unscrews the cap a quarter turn at a time, and uses its wrist force sensor, not a turn count, to know when the cap is free.

Start a projectHow it works
Lab automationThe full run, in physics simulation (MuJoCo), with models of the real arm, gripper and sensors.
Robots
Two UR5e arms with Robotiq 2F-85 grippers
Simulation
MuJoCo, with a 24-410 screw thread built to the real cap’s pitch and turns
Sensing
Overhead and wrist cameras, wrist force and torque
Learned parts
Optional: the unscrewing itself, learned from touch alone

Why it is hard

A screw cap is one of the most common things in a lab, and one of the hardest for a robot to be sure about.

  • A thread only lets go by turning

    Pulling a screw cap straight up does nothing. It has to turn about one and a half times, rising a little with each turn.

  • Fingers slip

    Each quarter turn of the wrist turns the cap a little less than a quarter, so counting turns does not tell you when it is off.

  • The last turn feels like the others

    Once the seal breaks, every quarter turn takes the same small twist, including the one where the cap comes free. Twist alone cannot tell.

  • The tube must not spin

    Breaking the seal takes far more twist than an ordinary grip on a glass tube can resist, so the holding arm needs a grip built for it.

How it works

The same steps you see in the video, in order.

  1. Step 1: Find and lift a capped tube

    The overhead camera finds the capped tubes. Arm A picks one from the side and holds it up above the middle of the table.

  2. Step 2: Lock the grip

    Arm A grips tighter, with the tube centred between its finger pads, and locks its fingers.

  3. Step 3: Grip the cap

    Arm B looks down with its wrist camera, comes down over the cap and grips it from above.

  4. Step 4: Quarter turn, then a lift test

    Turn the wrist 90°, then pull up a few millimetres. If the cap pulls back, it is still on: let go, turn back, grip again and repeat.

  5. Step 5: Cap off, tube to the rack

    When the pull is gone and the arm feels only the cap’s weight, Arm B carries the cap to its holder and Arm A puts the open tube in the rack.

Model or no model

Does unscrewing need a learned model?

Learned

  • An ACT policy that unscrews using only what the arm feels: force, torque and joint angles, no camera
  • Learned by watching the programmed version do the task
  • Trained with realistic sensor noise, so it never relies on readings a real sensor could not give

Programmed

  • The whole loop, written by hand: turn, lift test, grip again
  • Picking, holding and putting back the tube
  • A collision check before every move, including against the other arm

How we chose

We built the programmed loop first, as the baseline and as the teacher. The learned policy copies it using touch alone. It earns its extra cost on harder cases, such as caps screwed on by different amounts, where a fixed loop starts to struggle.

How we decide in general

Decisions that made it work

  • “Off” is felt, not counted

    A lift test after each quarter turn decides. Counting a fixed number of turns fails as soon as the fingers slip or the thread runs a little longer.

  • A second check before carrying

    Once the pull is gone, Arm B checks it is carrying only the cap’s weight. More would mean the tube is coming up with it, and the run stops safely instead.

  • A grip built for twist

    Arm A holds the tube centred between the pads, squeezes within the real gripper’s force range, and locks, as the real gripper does.

  • Physics checked against the hardware

    The seal, friction and thread follow published figures for 24 mm caps. A trial with the UR5e and 2F-85 models showed the lift test tells “on” from “off” with room to spare over a real wrist sensor’s error.

Tools we used

  • MuJoCo
  • UR5e
  • Robotiq 2F-85
  • Force/torque sensing
  • ACT
  • PyTorch
  • Python

Capabilities this draws on

Where it goes next

  • Caps screwed on by different amounts, where a learned policy can beat the fixed loop
  • Caps on the full source tubes, opened without spilling
  • Screwing caps back on to a set torque, as capping machines do

More use cases

Tell us what you want a robot arm to do.

Send a short description or a phone video of the task. We will reply with what we would simulate first, and whether a learned model belongs in it.

Email info@dodao.comRead our docs