Inverse Kinematics: Computing Joint Angles for Target End-Effector Poses

Inverse kinematics (IK) is the mathematical problem of computing the joint configuration of a robotic manipulator that places its end-effector at a specified position and orientation in Cartesian space. For a serial kinematic chain with n joints, IK maps from the 6-dimensional task space (3 translation + 3 rotation) to the n-dimensional joint space. When n > 6, the system is redundant and has infinitely many solutions; when n = 6, solutions are typically discrete; when n < 6, the system is under-actuated and cannot reach arbitrary poses.

The mathematical formulation involves inverting the forward kinematics function. Forward kinematics computes the end-effector pose T as a function of joint angles q: T = FK(q). IK seeks q such that FK(q) = T_target. For serial manipulators, FK is a product of homogeneous transformation matrices, one per joint. The inverse is non-trivial because FK is non-linear and may be non-injective (multiple q map to the same T).

Three solution strategies dominate: analytical methods that exploit specific kinematic structures to derive closed-form solutions (fast but only available for certain robot geometries), numerical methods that iteratively minimize the pose error using Jacobian-based optimization (general but slower), and sampling-based methods that search the joint space for valid configurations (most general but slowest). In practice, most modern robot arms have analytical IK solutions available through tools like IKFast, with numerical methods as fallback for singularity regions.

For teams building manipulation training pipelines, IK configuration directly affects data quality. During teleoperation, the IK solver determines how the operator's hand motions translate to robot joint motions. Solver inconsistencies between sessions — using different solution branches, different redundancy resolution strategies, or different joint limit handling — introduce artificial variation that the learning algorithm may interpret as meaningful.

Standard practice is to fix the IK solver and all its parameters (seed configuration, redundancy resolution method, joint limit margins) across all data collection sessions for a given dataset. This configuration should be documented and shipped with the dataset so downstream users can replicate the exact IK behavior if needed.

Claru documents the IK configuration used during data collection for every dataset, records both joint-space and end-effector-space action labels, and validates consistency between the two representations. This enables teams to train in either action space without concern about IK-related artifacts.

Claru records both joint-space and end-effector-space action labels in all demonstration datasets, with documented IK configurations. This dual representation supports flexible training paradigms while ensuring IK consistency across data collection sessions.

Inverse kinematics (IK) is the computational problem of determining the joint configuration of a robot that places its end-effector at a specified position and orientation in 3D space. Given a target pose — a 3D position (x, y, z) and a 3D orientation (roll, pitch, yaw or quaternion) — the IK solver computes the angles of each joint in the robot's kinematic chain that achieve that pose. For a typical 7-DoF manipulator like the Franka Panda, IK maps from 6D task space (3 translation + 3 rotation) to 7D joint space.

The forward kinematics problem — computing the end-effector pose given joint angles — is straightforward: multiply the transformation matrices for each link in the chain. Inverse kinematics is harder because the mapping from task space to joint space is generally non-linear, may have multiple solutions (redundancy), and may have no solution (the target is outside the reachable workspace). A 7-DoF arm has infinite IK solutions for most reachable poses because it has one degree of redundancy.

In the context of learned manipulation policies, IK plays two distinct roles. For policies that output end-effector space actions (position and orientation deltas), an IK solver converts these commands to joint-level commands that the robot controller executes. For policies that output joint-space actions directly, IK is implicit — the policy has learned the inverse mapping through training data. Understanding IK is essential for designing action spaces, interpreting policy behavior, and debugging deployment failures.

When a learned policy outputs end-effector actions (the dominant paradigm for cross-embodiment transfer), those actions must be converted to joint commands through IK. The choice of IK solver affects policy behavior: analytical solvers (IKFast) are fast but may produce discontinuous joint motions; numerical solvers (TRAC-IK, Levenberg-Marquardt) are smoother but slower; and optimization-based solvers can incorporate additional objectives like minimizing joint velocity or avoiding joint limits.

For teleoperation data collection — the primary source of manipulation training data — IK determines how the operator's hand motions are mapped to robot joint motions. The IK configuration (which solution branch to use, how to handle redundancy) affects the resulting demonstrations. Inconsistent IK settings between demonstrations can introduce spurious variation that confuses the learning algorithm. Standardizing the IK solver and its parameters across all data collection sessions is a critical data quality measure.

Some modern approaches bypass explicit IK entirely. Policies that output joint-space actions learn the inverse mapping implicitly from demonstrations. Diffusion Policy and ACT, when configured with joint-space action outputs, directly predict joint angle trajectories without any IK solver in the loop. This simplifies the deployment pipeline but couples the policy to a specific robot morphology, making cross-embodiment transfer harder.

The choice of action space in training data — joint-space vs. end-effector space — has significant implications for IK. Joint-space action data records the actual joint angles commanded at each timestep, capturing the specific IK solution used during data collection. End-effector space action data records the target pose, abstracting away the IK solution. Joint-space data is embodiment-specific but avoids IK-related artifacts at deployment; end-effector data enables cross-embodiment transfer but requires solving IK at deployment time.

For teams collecting demonstration data, recording both representations simultaneously is recommended: joint positions from the robot's encoders and end-effector pose from forward kinematics. This dual recording enables training policies in either action space without re-collecting data. It also provides a validation check — the forward kinematics of the recorded joint positions should match the recorded end-effector poses.

Claru's demonstration datasets include both joint-space and end-effector-space action labels, synchronized at the robot's native control frequency. This dual representation supports both embodiment-specific and cross-embodiment training paradigms, with consistent IK configurations documented for each dataset.