Connecting to sources…

How to build HopeJR (HOPEJr): the whole process

110 steps to build HopeJR (HOPEJr), from buying parts, printing and assembly to software, calibration and training, with the bill of materials, shop links and the offici…

No official price or BOM total: Arm/BOM.md lists parts and shop links without prices. A Hugging Face blog post (2026‑09‑25) puts the LeRobot-integrated arm and hand at around €500; the HF CEO's 2025‑05‑29 post says $3,000 for the full robot, but there is no order page

Bill of materials · Nuts & Bolts

PartQty
M2 · screw set · Amazon · hand1 set
M4 · nuts set · Amazon · arm1 set
65mm screws · Amazon · 2x upper shoulder, 2x armpit, 2x wrist6
50mm screws · Amazon · 1x hummeral1
45mm screws · Amazon · 1x hummeral1
35mm screws · Amazon · 1x hummeral1
30mm screws · Amazon · 1x hummeral, 2x lower shoulder3
25mm screws · Amazon · 1x upper elbow1
M6 · 35mm screws · Amazon · 2x elbow, 2x shoulder4
spacers set · Amazon · 2x elbow, 2x shoulder1 set

Bill of materials · Miscallenous

PartQty
PTFE tube · Amazon · hand1
Finger protectors · Amazon · 1x fingers5

Buy parts

  • Select the four listed motors

    Select the motors from the document's four-row motors table.

  • Buy one URT-1 board
  • Buy one Waveshare servo driver board

    Buy 1 Waveshare servo driver board for the hand. References: Amazon, AliExpress.

  • Buy the 12V 5A power supply

    Buy 1 power supply + DC adapter rated 12V - 5A for the arm. Reference: Amazon.

  • Buy the 5V 4A power supply

    Buy 1 power supply rated 5V - 4A for the hand. Reference: Amazon.

  • Buy two 3 m USB-C to USB-A cables

    Buy 2 USB C → USB A cables with a length of 3m for the arm + hand. Reference: Amazon.

  • Buy one M2 screw set

    Buy 1 set of M2 screws for the hand. Reference: Amazon.

  • Buy one M4 nuts set

    Buy 1 set of M4 nuts for the arm. Reference: Amazon.

  • Buy six 65 mm screws

    Buy 6 65mm screws for 2x upper shoulder, 2x armpit, and 2x wrist. Reference: Amazon.

  • Buy one 50 mm screw

    Buy 1 50mm screw for 1x hummeral. Reference: Amazon.

  • Buy one 45 mm screw

    Buy 1 45mm screw for 1x hummeral. Reference: Amazon.

  • Buy one 35 mm screw

    Buy 1 35mm screw for 1x hummeral. Reference: Amazon.

  • Buy three 30 mm screws

    Buy 3 30mm screws for 1x hummeral and 2x lower shoulder. Reference: Amazon.

  • Buy one 25 mm screw

    Buy 1 25mm screw for 1x upper elbow. Reference: Amazon.

  • Buy four M6 35 mm screws

    Buy 4 M6 35mm screws for 2x elbow and 2x shoulder. Reference: Amazon.

  • Buy one spacers set

    Buy 1 set of spacers for 2x elbow and 2x shoulder. Reference: Amazon.

  • Buy wrist bearing

    Buy 1 bearing sized 10mm x 70mm x 90mm for the wrist.

  • Buy shoulder bearing

    Buy 1 bearing sized 50mm x 65mm x 7mm for the shoulder.

  • Buy arm bearings

    Buy 18 bearings sized 20mm x 27mm x 4mm for the arm.

  • Buy short springs

    Buy 1 set of 10 springs sized 15mm x 0.3mm x 3mm; use 3 for the thumb and 2 for the fingers with the thumb.

  • Buy long springs

    Buy 2 sets of 10 springs sized 20mm x 0.3mm x 3mm; use 1 for the finger with the thumb.

  • Buy hand wire

    Buy 1 wire with 0.33 mm diameter and 63 LB test for the hand.

  • Buy PTFE tube

    Buy 1 PTFE tube for the hand.

  • Buy finger protectors

    Buy 5 finger protectors, using 1 for each finger.

  • Buy the M3 screw set

    Buy 1 set of M3 screws.

  • Buy the joint bearings

    Buy 10 bearings sized 6mm × 10mm × 3mm; use 2 for each upper-arm joint.

  • Buy forearm-size Velcro straps

    Buy 2 Velcro straps with a width of 20mm and a length matching the forearm diameter.

  • Buy shoulder-size Velcro straps

    Buy 2 Velcro straps with a width of 20mm and a length matching the shoulder diameter.

  • Buy half-torso-size Velcro straps

    Buy 2 Velcro straps with a width of 20mm and a length matching the half-torso diameter.

  • Buy encoder magnets

    Buy 23 magnets with a diameter of 3mm; use 1 for each joint.

  • Buy linear Hall sensors

    Buy 28 SS49E linear Hall sensors; use 1 for each hand and wrist joint and 2 for each upper-arm joint.

  • Buy 10 cm JST cables

    Buy 1 set of 10 three-pin JST cables measuring 10cm long.

  • Buy 15 cm JST cables

    Buy 1 set of 10 three-pin, same-direction JST cables measuring 15cm long.

  • Buy 20 cm JST cables

    Buy 2 sets totaling 20 three-pin, same-direction JST cables measuring 20cm long.

  • Buy 30 cm Three-Pin Wires

    Buy one set of 10 three-pin wires, each 30 cm long and oriented in the same direction.

  • Find deployment resources and released weights

    Open the project deployment page to find implementation notes, deployment notes, and released weights.

  • Obtain the glove hardware

    Obtain the ESP32-C3 board, firmware, KiCad PCB design, printable parts, URDF model, and interactive viewer from the Homunculus open hardware release. The glove materials cost under $20.

  • Consult the workflow documentation

    Use the LeRobot calibration and data-collection workflow in the Homunculus loco-manipulation documentation.

  • Set up HopeJR hardware

    Complete the linked HopeJR Hardware Setup before installing and configuring LeRobot.

Print and make

  • Obtain and print G1 modifications

    Obtain the print files from lerobot/unitree-g1-hardware-modifications and print the wrist adapters, head-camera holder, and backpack enclosure.

  • Print the glove parts

    Make the glove parts using the printable parts and STEP/STL geometry supplied by the open hardware release or Project Homunculus repository.

  • Locate the STL Files

    Find the STL files for HOPEJr and the teleoperation kit in 3dprint files/STL.

  • Prepare Arm and Exoskeleton Models

    Use the updated STEP/arm.step and STEP/exoskeleton.step files for the arm and teleoperation exoskeleton. The arm tolerances have been relaxed slightly.

  • Order the teleop PCB

    Add the Gerber file and select PCB assembly; on the next screen, paste the BOM and POS for the teleop device and connectors, then place the order.

Assemble

  • Add the Underactuated Finger Joint

    Add an underactuated degree of freedom at each finger's DIP joint so the hand can adapt to the object being gripped.

  • Lock and Thicken the Wires

    Lock the wires in place with screws and use thicker fishing wire to reduce shear and twist on the rolling contact joints.

  • Install PTFE Tube Tendons

    Use 3D-printer-filament PTFE tubes to drive the finger tendons.

  • Install the Opposable Thumb

    Install the thumb as an opposable structure using an actuated thumb-tracker approach.

  • Install the Finger Spools

    Add spools to linearize finger movement relative to motor angle, enabling faster action and 1:1 teleoperation.

  • Replace Threaded Inserts with Bolts

    Replace all threaded inserts in the arm with bolts.

  • Route Wires and Install the Arm Motor

    Route the wires behind the mantle, remove the passive tendon, and fit the bulkier STS8215 to handle the extra load.

  • Assemble the PCBs Without Soldering

    Use PCBs to connect the motors and simplify the teleoperation exoskeleton assembly. No soldering is required, including for the PCBs.

  • Build the dual joint exoskeleton

    Build the exoskeleton counterpart for every HopeJr robot joint, including the hand joints.

  • Fit joint angle encoders

    Use one diametric hall encoder for each hand joint and two for each arm joint to track angles more accurately.

  • Assemble without soldering

    Assemble the device without soldering by using the custom PCBs developed at HF.

  • Mount the OpenArm grippers

    Use the custom 3D-printed adapters to mount the OpenArm grippers onto the G1 wrists.

  • Mount the head camera

    Mount an additional camera on the G1's head with the printed holder, providing a stable egocentric view independent of the wrist cameras.

  • Install the backpack electronics enclosure

    Fit the printed enclosure on the G1's back to house the CAN-FD adapter and the additional gripper and camera cabling.

  • Fit the removable faceplate

    Fit the enclosure's removable faceplate so the interface remains accessible without dismantling the robot.

  • Build the head-depth vision pipeline

    Pass head-camera depth through frozen Theia-Tiny and a 16-frame temporal adapter to form a 128-dimensional visual summary.

  • Connect vision and body state to SONIC

    Feed the 128-dimensional visual summary and G1 body state through seven LoRA adapters into SONIC to produce joint targets.

  • Assemble sensors and controller

    Assemble the glove using the interactive 3D assembly guide, connecting 16 SS49E Hall-effect sensors, a multiplexer, and the custom ESP32-C3 board.

  • Build the Homunculus exoskeleton

    Extend the glove concept into the Homunculus exoskeleton for arm teleoperation. The exoskeleton measures joint motion directly and retargets it to the robot.

Software

  • Connect policy and whole-body controller

    In the LeRobot workflow on the Unitree G1, place a fast controller between the learned policy and the robot. Have the policy predict compact motion tokens and the controller convert them into dynamically feasible joint targets.

  • Configure π0.5 and SONIC

    Use the OpenHLM approach in which π0.5 takes language, camera observations, and robot state and predicts chunks of SONIC latent motion tokens. SONIC's decoder converts the tokens into joint targets that the low-level PD controller tracks.

  • Encode the reference into a motion token

    Convert the reference motion into SMPL format and pass it through SONIC's encoder to obtain a shared 64-D quantized motion token, which becomes the imitation-learning action target.

  • Decode and execute G1 motion

    Supply the motion token in the LeRobot integration; the decoder combines it with a short proprioceptive-state history to produce executable joint targets, which the controller converts into G1 motion.

  • Follow the G1 training and inference workflow

    Follow the reproducible training and deployment workflow in the G1 training and inference documentation.

  • Configure alternative whole-body controllers

    Control the G1 in LeRobot with NVIDIA GR00T-WholeBodyControl or Amazon FAR Holosoma, then compare them in the linked space.

  • Map G1 arm inverse kinematics

    Use G1 arm kinematics and inverse kinematics to map end-effector targets into joint targets.

  • Enable upper-body gravity compensation

    Enable gravity compensation for upper-body joints through the interface.

  • Install firmware and viewer

    Install the glove firmware and use the interactive viewer included with the open hardware release to view the model.

  • Install LeRobot and HopeJR dependencies

    Follow LeRobot's installation instructions, then install the HopeJR dependencies.

  • Integrate the exoskeleton with LeRobot

    Integrate the exoskeleton, angle encoders, and PCBs described above with LeRobot.

Calibrate and test

  • Identify HopeJR USB ports

    Run the port-finding script and note the displayed paths for the arm, hand, glove, and exoskeleton. The paths will be used for the --robot.port and --teleop.port parameters.

  • Start hand calibration

    Run the hand calibration script with the specified robot type, USB port, robot ID, and right-hand side. A calibration GUI will open, and calibration files will be saved in ~/.cache/huggingface/lerobot/calibration.

  • Calibrate the thumb joints

    In the calibration GUI, calibrate the thumb's CMC, MCP, PIP, and DIP joints individually and set each joint's range boundaries.

  • Calibrate the finger joints

    Individually calibrate the radial flexor, ulnar flexor, and PIP/DIP of the index, middle, ring, and pinky fingers. Set the ulnar and radial flexor ranges to the same size but with different offsets for symmetric movement.

  • Save the hand calibration values

    After setting the appropriate range boundaries for every joint, click “Save” to write the calibration values to the motors.

  • Calibrate the teleoperator glove

    Run the glove calibration command, then move the thumb and each finger through its full range of motion while recording positions. The completed calibration is saved to /Users/your_username/.cache/huggingface/lerobot/calibration/teleoperators/homunculus_glove/red.json.

  • Calibrate the robot arm

    Run the arm calibration command to open the GUI, then move every joint through its full range and adjust its minimum and maximum limits. Calibrate the shoulder pitch, yaw, and roll; elbow flex; and wrist pitch, yaw, and roll, then save the calibration.

  • Calibrate the arm teleoperator

    Run the calibration command for the black Homunculus arm teleoperator connected to /dev/tty.usbmodem11201.

  • Move joints and record positions

    Move all joints through their entire range of motion, then press ENTER to stop recording. The calibration is saved to /Users/your_username/.cache/huggingface/lerobot/calibration/teleoperators/homunculus_arm/black.json.

  • Validate reliable dodging in simulation

    Evaluate the training in simulation. PPO with reward alone did not learn to dodge reliably, while the visual signal was key.

  • Calibrate fingers and map hand

    Use the glove as the HomunculusGlove teleoperator in LeRobot, read the serial sensor stream, and complete per-finger range-of-motion calibration. After calibration, map the glove’s motion to the HOPE-Jr hand.

Data and training

  • Collect whole-body teleoperation demonstrations

    Collect robot demonstrations through whole-body teleoperation for training loco-manipulation behavior.

  • Record 64-D SONIC latent actions

    Record the 64-D latent action representation produced by SONIC's encoder instead of the robot's joint targets.

  • Train the VLA on latent actions

    Train the vision-language-action model with the recorded SONIC latent action representation so that it predicts this representation.

  • Execute actions with SONIC decoder

    Pass the trained policy's output to SONIC's decoder for execution. The decoder produces joint targets that the robot's low-level PD controller tracks.

  • Load the can pick-and-place dataset

    Use nepyope/can_clean_final to train a can pick-and-place policy.

  • Collect whole-body teleoperation demonstrations

    Record roughly 100 episodes at 50 fps, totaling about 71 minutes. Each episode uses the instruction “Bring the can to the white table.”

  • Record camera views and robot state

    Record three 480×640 camera streams at every timestep. The 31-dimensional state contains 29 G1 joint positions and two gripper states.

  • Define the SONIC policy action

    Use a 66-dimensional action: a 64-dimensional SONIC latent plus two gripper commands.

  • Fine-tune the pi05 vision-language-action model

    Fine-tune lerobot/pi05_base for 12,000 steps on 4×H100 GPUs.

  • Replicate the ViBE and ORCS dodge experiment

    Train a LoRA adapter on SONIC in mjlab with PPO so the robot dodges thrown balls using head-camera depth data and its body state.

  • Add ball position and velocity prediction

    During training, use an auxiliary objective to ask the percept to predict the ball’s position and velocity.

  • Visualize G1 data and motion

    Play back the G1 URDF in the LeRobot visualizer to inspect robot motion visually.

  • Read the full G1 proprioceptive state

    Read joint positions and velocities, IMU data, estimated joint torques, and motor temperatures from the interface.

  • Record teleoperation datasets

    Use the Homunculus glove to directly measure finger-joint motion for recording dexterous-manipulation datasets and teleoperation data.

  • Use the public G1 dataset

    The public nepyope/unitree_box_move_blue_full example contains 550 episodes and roughly 475k frames of G1 teleoperation data.

  • Separate arm and hand sessions

    Run arm teleoperation and hand teleoperation in separate shell sessions because of global variable conflicts in the Feetech middleware.

  • Teleoperate the right hand

    In one shell session, start teleoperation for the blue hope_jr_hand robot and red right-hand Homunculus glove, with data display enabled at 30 FPS.

  • Teleoperate the arm

    In another shell session, start teleoperation for the white hope_jr_arm robot and black Homunculus arm teleoperator, with data display enabled at 30 FPS.

  • Record a Hope-JR hand dataset

    Record one 5-second episode with the right Hope-JR hand and homunculus glove, using a 640×480 OpenCV camera at 30 fps. The private dataset is pushed to the Hub with streaming encoding enabled; recording, replay, and training are experimental.

  • Replay a recorded Hope-JR episode

    Replay episode 0 from <USER>/hand_record_test_with_camera using the right Hope-JR hand. Replay with Hope-JR is experimental.

  • Train an ACT Hope-JR policy

    Train an ACT policy on <USER>/hand_record_test_with_video_data using the MPS device, with tracking enabled. Save the output to outputs/train/hopejr_hand and set the policy repository to <USER>/hand_test_policy; training with Hope-JR is experimental.

  • Evaluate the HopeJR Hand Policy

    Record 10 evaluation episodes for the right HopeJR hand using the pretrained policy and a 640×480 OpenCV camera at 30 fps.

  • Record data and train HopeJR

    Use LeRobot to record data, train, and perform inference on HopeJR.

Use it

  • Dodge using live head-camera depth

    During rollout, the G1 directly uses head-camera depth data to dodge in real time.

  • Use the HOPE-Jr humanoid

    Use these devices to drive HOPE-Jr. HOPE-Jr is a full-size humanoid with dexterous, tendon-driven hands built from 3D-printed parts and off-the-shelf servos; the LeRobot-integrated arm and hand alone cost around €500.

Common problems

  • BOM leaves a tendon motor unspecified, blocking correct part selection #1
  • The PCB connector is unavailable and editable KiCad files are missing #11
  • The STEP file cannot be opened in FreeCAD #12
  • Missing license terms block legal reuse, modification, and sharing #3 #6

Parts

  • Hand: 16 Feetech SCS0009 micro servos; 0.33 mm 63 lb fishing line as tendons, PTFE tubes as guides, tension springs for return; an underactuated DIP joint and an opposable thumb
  • Arm, 7 joints (shoulder pitch/yaw/roll, elbow, wrist pitch/yaw/roll): SM8512BL ×1 (shoulder pitch), STS3250 ×2 (shoulder yaw, elbow), STS3215 12V ×4
  • URT-1 board (arm) and a Waveshare servo driver board (hand), 12V 5A and 5V 4A supplies, two 3 m USB‑C to USB‑A cables
  • Bearings: 70×90×10 mm (wrist), 50×65×7 mm (shoulder), 20×27×4 mm ×18; M2/M4/M6 screws and M6 spacers
  • Homunculus exoskeleton and glove: an ESP32‑C3‑WROOM‑02 + CD74HC4067 multiplexer PCB and 7/12-port connector boards (JLCPCB assembly), 28 SS49E Hall sensors, 23 3 mm diametric magnets, JST leads and Velcro straps
  • Humanoid/ drafts: SolidWorks 2018 CAD, a right-leg and pelvis URDF, legs moving to Robstride RS02 motors (2026‑04 test part)

Tools

  • A 3D printer (STLs for left and right arms, hand and exoskeleton are in the repo)
  • A JLCPCB account to order the assembled PCBs from the Gerber, BOM and placement files
  • The README says threaded inserts were all replaced by bolts, so with the custom PCBs no soldering is needed

Skills

  • Python and the LeRobot CLI (lerobot-calibrate, -teleoperate, -record, -train)
  • Joint-by-joint calibration: finger ranges set in a GUI, glove and exoskeleton moved through their full range
  • Flashing the exoskeleton and glove Arduino firmware; training policies such as ACT

Software

  • LeRobot (Apache‑2.0): pip install -e ".[hopejr]", robot types hope_jr_arm / hope_jr_hand, teleop types homunculus_arm / homunculus_glove
  • Arm/code: exoskeleton and glove Arduino firmware (exo.cs, glove.cs; 16 Hall channels read through a multiplexer)
  • Humanoid/URDF: right-leg and pelvis simulation model

Documents