Connecting to sources…

K2 (open-source 3D-printed 12-DoF biped) · A small 3D-printed biped with no upper body: 12 Feetech STS3215 servos (per leg: hip pitch, roll and yaw, knee, ankle pitch and roll), 0.426 m tall and 1.40 kg, with a reinforcement-learned gait running on a XIAO ESP32-S3 microcontroller

A small 3D-printed biped with no upper body: 12 Feetech STS3215 servos (per leg: hip pitch, roll and yaw, knee, ankle pitch and roll), 0.426 m tall and 1.40 kg…

Maker
Kapalik Khanal open-source project
Form
A small 3D-printed biped with no upper body: 12 Feetech STS3215 servos (per leg: hip pitch, roll and yaw, knee, ankle pitch and roll), 0.426 m tall and 1.40 kg, with a reinforcement-learned gait running on a XIAO ESP32-S3 microcontroller
For
Researchers and reinforcement-learning hobbyists; the README calls K2 a research platform, and the project site says it is not yet a complete kit (no assembly guide, print settings, wiring diagram or editable CAD)
Sensors
An LSM6DS3 IMU over I²C plus joint position and velocity from the 12 servos; the README says sensing is proprioceptive only and nothing on board handles perception or planning. Seeed's docs say the XIAO ESP32-S3 Sense board carries an OV5640 camera and a digital microphone, but the K2 docs do not use them
Connectivity and app
The ESP32-S3 runs its own Wi‑Fi access point, K2-Robot, controlled from a phone browser at 192.168.4.1 or from an Xbox controller over BLE; on a Raspberry Pi 5, keyboard.py or joystick.py runs the same policy. No app and no cloud service
Subscription
Open-source project, no subscription
Openness
The GitHub API reports MIT (“Copyright (c) 2026 K2 contributors”), covering the ESP32-S3 firmware, the Raspberry Pi runtime, the mjlab training code, the shipped ONNX policy, the MJCF/URDF digital twin and STLs for 18 structural parts; the README says dependencies keep their own licences. The project site says the repo has meshes only, with no editable CAD (.f3d/.step)
Privacy
The docs use no camera or microphone, and control is entirely local. The project site warns that the ESP32 web control's access point uses a documented default password and its local control endpoints are unauthenticated: use it on a trusted network, change the password for public demos, and never expose it to the internet.
Not stated
Battery type, capacity and runtime, assembly time, print settings and filament, editable CAD, the creator's location, kit price

Vendor claims

  • The learned gait runs entirely on a microcontroller: a 55,088-parameter policy trained with PPO in MuJoCo runs at 50 Hz on a Seeed XIAO ESP32-S3, with no Linux computer or GPU on board
  • One policy covers hold, march, forward/backward walking, turning and lateral stepping, and the same policy also runs on a Raspberry Pi 5
  • A $320.52 bill of materials; the digital twin can be driven from the keyboard in MuJoCo with no robot or GPU

Parts

  • ESP32-S3R8 · Runs K2 control, sensing, and wireless communications
  • Raspberry Pi 5 · Runs the policy and controls robot servos
  • XIAO ESP32-S3 · Serves as K2 controller and connects sensors
  • XIAO ESP32-S3 Sense · Runs inference and handles IMU, servo telemetry, and joint commands
  • OV2640 · Captures 1600×1200 fixed-focus images
  • OV3660 · Captures 2048×1536 fixed-focus images
  • OV5640 · Captures 2592×1944 autofocus images
  • LSM6DS3 · Measures robot body orientation and acceleration
  • ST3215 · Drives the robot's twelve joints
  • STS3215 · Drives 12 joints and reports position and velocity
  • CH343 · Bridges controller to servo serial bus
  • U.FL-R-SMT1 · Connects the onboard U.FL antenna
Parts and sources

Sources