Inspiration

Hands do two very different jobs: some tasks need force, others need finesse. Arthritis, Parkinson's, stroke and simple weakness take away both. The most common example is the child-resistant pill bottle, which demands a hard squeeze and a twist at the same time. Most robot grippers are no better: they close to a fixed position, so they're either too weak for a stuck cap or strong enough to crush a piece of fruit. We wanted one hand that adjusts its grip by feel.

What it does

Tend is an assistive robot arm with a pressure-sensing fingertip. It runs the same loop for every task: see (a wrist camera plus a second camera watching the workspace), reach, feel (close until the fingertip reads a target pressure), then act and check.

  • Opening child-resistant pill bottles (our main demo). Set the bottle down. Tend lifts it by the cap into a clip, a servo clamp centres and holds it, and the gripper squeezes the cap firmly. The arm presses down to release the child lock, then twists the cap off while watching for slip.
  • Handling soft and fragile things. The same grip loop with a lower pressure target picks up fruit, bread or a paper cup without crushing it.
  • Checking its own work. If the pressure drops or shifts mid-move, the object is slipping: it regrips tighter and retries. After a task it confirms by touch and camera that it worked.

How we built it

  • Arm: a 5-axis printed arm (Dynamixel XL330 servos, printed 4:1 planetary shoulder) with an SO-101 gripper, controlled through Hugging Face LeRobot. We reconstructed the true joint positions from the parts' mating faces and hole patterns, so the software model matches the physical arm.
  • Vision: a wrist camera to line up the gripper, and a second camera watching the workspace to detect objects and confirm a bottle is seated in the clip.
  • Touch: a TouchTronix SensX25 5×5 pressure pad on the gripper finger. We wrote a driver, diagnostics, a live heatmap and a network server so several programs can share the sensor.
  • Fixture: a printed clip and a goBILDA servo clamp with coupled jaws, run by our ESP32 hub firmware.
  • Simulation: a MuJoCo digital twin of our arm and gripper STLs with a physically modelled child-resistant cap (thread torque, push-to-unlock spring, randomized per run). The same control code runs on the robot and in the sim. We trained a PPO reach policy (median 3.8 mm error) and exported a URDF for ROS.

Challenges we ran into

  • The cap sits 12 mm off the wrist axis. Turning only the wrist swings the cap in a circle, so the jaws scrub instead of unscrewing. We solve inverse kinematics every 3° so the whole arm orbits the cap's own centre.
  • Grip strength. The sim showed our original gripper servo can't turn caps tighter than about 0.1 N·m (3 / 16 opened); a stronger servo opened 10 / 16.
  • Pushing and twisting at once. Pressing down through the jaws uses the same friction the twist needs.

Accomplishments that we're proud of

  • Grip by feel, with slip detection, on a commercial tactile array.
  • A simulator that changed our hardware decisions before we printed parts.
  • The finished arm with all the modification to solve small but challenging issues we fixed and tested for repeatedly.

What we learned

  • Yue: [designing an arm whose pivots line up with real hole patterns]
  • Julian: [taking printed parts from CAD to a working assembly in a day]
  • Alan: [building a control framework that runs the same on the robot and in simulation]
  • Henri: [streaming a tactile array and turning pressure into grip decisions]

What's next for Tend

  • Finish the full pick → place → open run on hardware (in sim the clamp and placement work, but the bottle often lands a few mm off centre).
  • More tasks on the same loop: handing items to a person, opening jars, sorting a weekly pill organizer.
  • Voice start ("Tend, open my morning bottle") and medication reminders.
  • Test with occupational therapists and people with arthritis.

Credit to

5DOF_Robot_Arm (danslovich), SO-101 (TheRobotStudio), LeRobot (Hugging Face), sensx_python (TouchTronix), MuJoCo (Google DeepMind), Stable-Baselines3, OpenCV.

The SensX25 is what lets one gripper be both strong and gentle. Tend closes on measured pressure instead of a fixed position, with a high target for a pill cap and a low one for soft items. It detects slip during a twist from changes in total load and the pressure centroid, and it confirms a task worked by checking the pad still feels the object. We wrote a probing and diagnostics tool, a live heatmap, a network server so several programs can share one sensor, and a virtual SensX25 that speaks the exact serial protocol for testing without hardware.

Built With

  • 3d-printing
  • autodesk-fusion-360
  • dynamixel
  • esp32
  • gymnasium
  • lerobot
  • lerobot-(hugging-face)
  • mujoco
  • mujoco-(google-deepmind)
  • opencv
  • pyserial
  • python
  • raspberry-pi
  • sensx-python-(touchtronix)
  • so-101-(therobotstudio)
  • solidworks
  • stable-baselines3
  • touchtronix-sensx25
Share this project:

Updates

Submission history