Skip to content

EmbodiK

High-performance prioritized numerical inverse kinematics for cross-embodiment robotics and VLA/AI applications

EmbodiK is a modern C++ prioritized numerical IK library with Python bindings for robot bring-up, teleoperation, retargeting, and whole-body IK examples. It is built on Pinocchio, exposed to Python through Nanobind, and keeps task hierarchy, constraints, and recovery policy in the solver so examples can stay focused on targets and visualization.

Design Principles

Cross-EmbodimentApply one IK policy across arms, humanoids, quadrupeds, and mobile bases.
Solver IntelligenceKeep priority, recovery, diagnostics, and constraint policy in the solver.
Constraint-AwareApply declared joint, collision, CoM, and supported kinematic-contact constraints.
ScalableUse the same solver surface for demos, stress tests, and policy rollout.

See the 0.24.0 release scope for functional support, experimental APIs and the limits of recovery and real-time claims.

GPU WBC: 1,024 Independent Worlds

The experimental model-derived GPU path uses Newton for batched kinematics and Warp directional SRINV for prioritized velocity IK. A single public example selects Panda, AI Worker, or G1 and derives the active joints and task shape from the loaded model. Its viewer renders 512 articulated robots by default using shared per-link mesh instances. Four colored world bands follow distinct motion families, with independent phases and speeds inside each band. Pass --worlds 1024 to reproduce the full reference scale.

Model Active DoF 6D tasks / world Solve-only p50
Franka Panda 7 1 0.906 ms
ROBOTIS AI Worker SG2 15 2 4.854 ms
Unitree G1 29 4 20.295 ms

Profile: 1,024 CUDA-resident worlds, two solver iterations, 50 measured steps after 20 warm-up steps, CUDA-capable NVIDIA GPU with approximately 24 GB of device memory. Target generation, visualization, collision, and host publication are excluded. See GPU WBC for the full capability matrix and Parallel GPU WBC to reproduce it.

✨ Features

  • ⚙️ Fast C++ core: Eigen-based IK routines with Python bindings and numpy support.
  • 🎯 Prioritized task hierarchy: Register frame, posture, relative-pose, and CoM objectives with explicit priorities.
  • 🛡️ Solver-owned robustness: Elastic limits, adaptive dt, auto task layout, weighted fallback, stall recovery, and collision guards — in C++.
  • ⚡ Tunable collision stack: Speed / Balanced / Precise presets, conservative bounds, and post-step guards — often ~10–50× faster collision steps vs naive full scans.
  • 🔒 Hard-constraint handling: Joint limits, collision constraints, CoM support polygons, contact, and relative-pose checks stay in C++.
  • 📈 Diagnostics: Timing, condition numbers, recovery stage, task scaling, and solver status reporting.
  • 🎮 Examples and visualization: Viser and mjviser demos for Panda, AI Worker, RB-Y1, Unitree G1, and Spot workflows.
  • ⚡ Model-derived GPU WBC: Newton/Warp CUDA execution for one interactive robot or thousands of independent worlds, without robot-family dimension constants.

🚀 Quick Start

Install EmbodiK, run a maintained example, then adapt the registered-task API pattern from that script:

python -m pip install --only-binary=:all: "embodik[examples]"
python -c "import embodik; print(embodik.__version__)"
embodik-examples --copy
cd embodik_examples
python 01_basic_ik_simple.py

See the Quickstart and Examples pages for the current KinematicsSolver, add_frame_task(), and solve_position_step() workflow.

📦 Installation

See the Installation Guide for wheel, source-build, and troubleshooting instructions.

python -m pip install --only-binary=:all: embodik

📚 Documentation

Learn

  • Quickstart — Build a small prioritized IK example with registered tasks.
  • Examples — Run maintained public examples and clone-only development demos.
  • Guides overview — Pick a reading path (robustness, collision, GPU, transforms).

Configure the solver

  • Solver Robustness & Recovery — Adaptive dt, elastic limits, auto layout, weighted fallback, stall recovery.
  • Acceleration Solver — Fixed-base acceleration eSNS, compatible state boxes, effort/contact rows, and collision certification boundaries.
  • Collision Constraints & Tuning — Tuning presets, safety layers, and performance vs naive full-scan checks.
  • GPU WBC — Newton/Warp setup, model compatibility, constraints, batching, and measured performance.

Reference

🎬 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

📄 License

Apache License 2.0 - see the LICENSE file for details. Source: robodreamer/embodik

Copyright (c) 2026 Andy Park andypark.purdue@gmail.com