Pink

Pink supplies Python differential inverse kinematics for articulated robot models using Pinocchio and quadratic programming. It combines weighted task errors with configured limits and optional barriers. This is model/solver software, not a physical robot driver or a promise that every requested task is feasible. [1] [5]

The old stephane-caron/pink repository resolves to pink-kinematics/pink. This entry preserves the current project identity rather than counting that rename as another implementation. [6] The source and installation documentation identify the package name as pin-pink on PyPI and pink on conda-forge. [1] [3]

The installation guide recommends conda install -c conda-forge pink and offers pip install pin-pink. Its source fallback requires separately installed eigenpy/Pinocchio before installing without dependencies. The README's concrete example loads a model with upkie_description, selects quadprog and uses a 0.006-second integration timestep. [1] Those are upstream example choices, not verified compatibility or recommended values for every robot. Match the actual model, libraries and backend; no package or solver was installed here. [3]

For a bounded first model task, the README creates FrameTask/PostureTask objects, sets targets and assigns costs. Frame names such as base and left_contact belong to the selected robot model, not arbitrary strings guaranteed to exist. Position/orientation costs normalize different units and trade off competing tasks; a weighted objective is not an independent demonstration that every target can be achieved. Pink is a local method rather than a global IK solver and can stop at a local optimum constrained by configuration limits. A stationary configuration therefore does not prove a reached target or the absence of another solution. [1]

The inspected solve_ik.py constructs objective and inequality terms from tasks, limits, barriers and the integration timestep. Match the model's configuration/velocity limits and solver result handling before applying a model update. A computed velocity or integrated configuration does not establish motor transport, feedback accuracy or collision safety on hardware. [5]

The project is Apache-2.0; external model meshes, Pinocchio and solver dependencies retain their own terms. Begin with one matched model and task configuration, then inspect solver behavior and limits before adding a robot interface. No IK solve, model integration, numerical comparison, trajectory execution or physical-control test was performed. [1] [2] [5]