embodik 0.20.9

High-performance inverse kinematics solver optimized for cross-embodiment VLA/AI applications

EmbodiK

EmbodiK is a high-performance inverse kinematics library for cross-embodiment robotics and VLA/AI applications. It pairs a C++ core with Python bindings, exposes robot-model utilities without requiring the Python pin package at runtime, and includes interactive examples for collision-aware IK, CoM constraints, teleop, GPU batch solving, and dual-arm coordination.

Overview

EmbodiK is designed for bringing up IK behavior across different robot bodies without rewriting the solver stack for each model. The public examples focus on a practical path:

• start with the smallest fixed-base IK loop.
• add collision-aware behavior and visualization.
• connect the same stepping IK pattern to teleop input.
• use richer clone-only examples for development and stress testing.

The detailed installation notes, API reference, and example walkthroughs live in the official documentation:

https://robodreamer.github.io/embodik/

Quick Start

Fastest path for most users is the wheel-only PyPI install inside a virtual environment:

python -m venv .venv
source .venv/bin/activate
python -m pip install -U pip
python -m pip install --only-binary=:all: embodik
python -c "import embodik; print(embodik.__version__)"

If that import works, the core package is installed.

If pip cannot find a compatible wheel, use the one-shot source installer for your platform from the Installation Guide. It creates the venv, installs native dependencies, builds EmbodiK, and runs an import smoke test.

To run copied examples from the same venv:

python -m pip install "embodik[examples]"
embodik-examples --copy
cd embodik_examples
python 01_basic_ik_simple.py

Published repaired wheels do not require the Python pin package at runtime. Source builds use pin or a system Pinocchio install as the native library provider; keep that provider installed in the environment used to import EmbodiK.

Examples

The pip-facing examples are intentionally split by purpose:

Script	Purpose
01_basic_ik_simple.py	Minimal fixed-base IK bring-up for a robot preset or new URDF.
02_collision_aware_IK.py	Collision-aware IK behavior demo and advanced tuning surface.
03_teleop_ik.py	Small adapter showing how teleop input drives the same IK step.
08_com_constraint_example.py	CoM support-polygon constraint visualization.
09_dual_arm_ects.py	Dual-arm ECTS and orthogonal coordination modes.
12_bimanual_whole_body_ik.py	Bimanual whole-body teleop, defaulting to AI Worker and optionally supporting RB-Y1, with CoM and collision handling.
13_unitree_g1_retargeting_ik.py	Unitree G1 whole-body retargeting IK with CoM and optional collision handling.
14_spot_full_body_ik_viser.py	Spot full-body IK in regular Viser with arm+torso, torso-only, full-body, and two-stage modes.
15_spot_locomanip_mjviser.py	Spot locomanipulation ONNX policy rollout in MuJoCo through mjviser.

Run them from a copied example directory:

python 02_collision_aware_IK.py
python 03_teleop_ik.py
python 12_bimanual_whole_body_ik.py
python 13_unitree_g1_retargeting_ik.py
python 14_spot_full_body_ik_viser.py

The regular Viser Spot full-body IK example uses the standard example dependencies and includes a bundled public Spot-with-arm URDF. The MuJoCo/mjviser locomanipulation example needs the optional mjviser stack and includes a bundled public MuJoCo Menagerie Spot-with-arm MJCF scene. mjviser is the MuJoCo-backed web viewer environment for policy rollout and interactive simulation. mjviser-teleop is the same viewer stack plus the optional Seer/xvisio controller dependencies, so use it only when running --enable-teleop:

python -m pip install "embodik[mjviser]"
embodik-examples --copy
cd embodik_examples
python 15_spot_locomanip_mjviser.py --policy locomanip
python 15_spot_locomanip_mjviser.py --policy locomanip-stationary
# add the optional Seer controller extra when needed:
python -m pip install "embodik[mjviser,teleop]"
python 15_spot_locomanip_mjviser.py --enable-teleop --policy locomanip

From a repository checkout, run the same example from the repository root with the matching Pixi environment:

pixi run -e mjviser spot-locomanip-mjviser --policy locomanip
pixi run -e mjviser-teleop spot-locomanip-mjviser --enable-teleop --policy locomanip

Use a single Pixi environment per run: mjviser for browser GUI control and mjviser-teleop for browser GUI plus Seer controller input. Do not add mjviser after python; it is a Pixi environment name, not a Python module argument. Passing --enable-teleop connects the optional controller; the in-app Enable teleop box starts enabled when the controller connects. The locomanipulation app also has solver preference sliders. Base assist controls how much the solver uses x/y/yaw locomotion while tracking the gripper target: lower values keep more motion in the arm, and higher values let the base help earlier. Arm recovery bias increases the arm posture return while condition-number protection is active, which helps avoid fully stretched arm configurations during loco-manipulation teleop. The default values are tuned for teleop: the base helps on reachable x/y/yaw nudges, and arm recovery stays active without making the high-condition-number solve overly stiff. The mjviser loop rate-limits ONNX policy inference and background IK requests to 50 Hz by default while MuJoCo simulation and rendering continue stepping at the model/viewer rate. The IK worker keeps only the newest pending request so collision-constrained solves cannot build a backlog and starve the sim thread. When collision avoidance is enabled, the default checks the 3 closest active collision pairs with balanced speed/accuracy tuning. Clone-based Spot IK changes can be headlessly checked with:

pixi run -e mjviser-teleop python scripts/spot_locomanip_ik_hardening.py

Most examples default to the Panda preset. Use --robot <key> when a script supports alternate robot presets. See the Examples Guide for the full catalog, helper conventions, and clone-only development examples.

Preview

Franka Panda collision-free IK

ROBOTIS AI Worker constraint teleop

RB-Y1 bimanual whole-body IK

Unitree G1 retargeting IK

Spot full-body IK

Spot locomanipulation mjviser

Core Capabilities

• C++ IK core with Nanobind Python bindings.
• Hierarchical velocity IK tasks for frames, posture, CoM, and dual-arm coordination.
• Joint-limit, self-collision, and CoM support-polygon constraints.
• Solver diagnostics for timing, task scaling, and Jacobian condition-number logging.
• Lie-group-aware configuration operations for floating-base, quaternion, and continuous joints.
• Native Pinocchio-backed robot model utilities exposed through EmbodiK bindings.
• Optional Viser visualization for interactive IK demos.
• Experimental GPU batch IK and collision tooling for high-throughput research workflows.

Documentation

• Installation - platform setup, source builds, and troubleshooting.
• Quickstart - first IK calls and solver concepts.
• Working with Transforms - transform helpers and SE(3) operations.
• Examples - public scripts and development-only demos.
• API Reference - Python API generated from docstrings.
• GPU Solvers - FI-PeSNS and PPH-SNS batch solver notes.
• Development - local build, tests, and contributor workflow.

Development

Use Pixi from a repository clone:

pixi run build
pixi run test
pixi run docs-build

Run examples from the clone:

pixi run python examples/01_basic_ik_simple.py
pixi run python examples/02_collision_aware_IK.py
pixi run python examples/03_teleop_ik.py

Clone-only advanced surfaces live under dev_examples/ and are not copied by embodik-examples --copy.

Repository Layout

embodik/
|-- README.md
|-- cpp_core/
|   |-- include/embodik/
|   `-- src/
|-- python_bindings/
|   `-- src/
|-- python/embodik/
|-- examples/
|-- dev_examples/
|-- docs/
|-- scripts/
`-- test/

Star History

License

EmbodiK is released under the Apache License 2.0. See LICENSE for details. Binary wheels may bundle permissively licensed native dependencies; see THIRD_PARTY_NOTICES.md.

Developer: Andy Park [e-mail]