Open source humanoid robot. 3D printed structure, 30 RobStride quasi-direct-drive actuators, NVIDIA Jetson AGX Thor, Unitree G1 kinematics as reference. Still in design, not built yet: shared now so it can be built together.
CAD on Onshape · BOM · Electrical scheme · Printable STL · URDF · Simulation · Design notes · River Family
▶ Video in 4K · Download the whole folder · Start here to continue the project, also with an AI
- height ≈ 1.40 m, purchased mass 33 kg, finished estimate ≈ 45 kg
- 30 DoF: legs 6 + 6, arms 7 + 7, waist 2, neck 2
- ankle: two motors in the shin, two pushrods, offset gimbal — pitch is the sum, roll the difference
- actuators, all RobStride QDD: RS04 ×6 hips and knees, RS03 ×3 waist roll and hip yaw¹, RS06 ×11 ankles, waist yaw, shoulders, elbows, RS00 ×8 shoulder yaw and wrists, RS05 ×2 neck
- power: 48 V 13S2P Li-ion 421 Wh, 24 V rail for computer and hands, hardware e-stop with precharge
- compute: Jetson AGX Thor, 5 CAN buses at 1 Mbps, RealSense D436, pelvis IMU
- hands: Inspire RH56DFX, an own 15-DoF hand is in development (docs/hand-design.md)
- structure: PA-CF / PPA-CF, Bambu Lab H2C, 0.4 nozzle
- cost: ≈ EUR 30 000 all new; ≈ EUR 10 300 if Thor and hands are already owned
¹ hip yaw: the CAD currently mounts RS06, the BOM RS03. Decided by simulation before ordering the legs.
| Phase | Content | EUR incl. VAT |
|---|---|---|
| 1 | right arm, right hand, one ankle kit, battery, safety chain, torso on a table | 4 009 |
| 2 | legs and locomotion | 3 998 |
| 3 | left arm, waist, neck, audio, mobile | 2 258 |
Filter the Phase column in bom/bom.xlsx. Every row has supplier, link, mass and notes. Thor and hands are counted at zero because they are owned; reference prices are in the notes.
| Folder | Content |
|---|---|
| bom/ | bom.xlsx 108 active rows, bom.csv, build_bom.py generator |
| electrical/ | connection scheme, written and drawn |
| cad/ | Onshape link, printable STL of the structure, first 3MF files, RobStride FeatureScript |
| cad/bonegen/ | generated bones: a bone is grown by topology optimisation between the screw holes of two motors, clear of everything that moves. Thigh pilot, plan for the whole robot |
| sim/ | URDF with meshes, mirror pipeline, ankle transmission map |
| sim/isaaclab/ | walking: trained policy, Isaac Lab task code, how it was reached round by round, traps, results, next steps |
| docs/ | why RobStride and not Encos or CubeMars, actuator tables, hand design, references, AI handoffs and working memory |
| AGENTS.md | entry point to continue the project, for people and AI agents |
What is printed is in cad/stl/. What is bought — motors, hands, battery, computer, camera, pushrods, screws — is in the BOM.
The right side and the centre are modelled in Onshape; sim/mirror_urdf.py mirrors them into the full 30-joint URDF with masses, limits and collisions: sim/urdf/primo.urdf.
Walking is trained in Isaac Sim 5.1 / Isaac Lab 2.3.2 with RSL-RL PPO, on rough terrain with stairs of 5–23 cm: 12 leg joints + waist roll, mean knee flexion 36°, about 87 % of a 1000-step episode survived. Everything is in sim/isaaclab/: the policy kneehard_model_2999.pt with the exact parameters it was trained with, the task code, the diagnostic scripts, the story round by round, the traps, the measured results and the next steps.
It is at 2:06 of the video. Clips: KneeHard, the chosen policy · KneeTorque.
All RobStride, after comparing Encos (the motors of the Asimov robot), CubeMars, Steadywin, MyActuator, ZeroErr and Dynamixel. Encos is smaller and lighter at the same torque, but its 25–36:1 reduction gives up backdrivability and force control for good, worst on the ankle, and it is sold on quotation without public CAD. The comparison and the reasons: docs/motor-selection.md, every number in docs/actuator-tables.md. Standard modules against custom actuators whose housing is the bone, and how a later version could get there: docs/actuator-integration.md.
- nothing built yet: phase 1 is the first bench
- before ordering: battery shipping to Italy and BMS limits, real ankle crank geometry, hip yaw motor — measure its torque while turning in simulation, RS06 36 Nm or RS03 60 Nm (how)
- walking policy: no pushes, no domain randomisation and arms held still so far; not yet ready for the real robot
- ankle torque map is derived from the CAD export and still to be verified
- e-stop does not cut hand power: decide whether it should
- STL are exported from the design state, not print-validated
- the CAD femur touches the shank assembly from about 123° of knee flexion (URDF limit 165°): decide the real joint limits (cad/bonegen/)
- generated bones are a pilot with assumed loads: next, loads logged from the walking policy and 2 mm resolution
Printables · Thingiverse · Cults · MakerOnline · YouTube · riverfamily.art
CC0 1.0. One exception: the task code in sim/isaaclab/code/ derives from Isaac Lab and keeps its BSD-3-Clause licence. Vendor files are not redistributed as CAD: motor and hand STEP files, manuals and the Unitree URDF are linked from docs/references.md.