Joint-Space Control: Direct Command of Robot Joint Configurations

Joint-space control is a robot control paradigm that operates directly on the robot's joint-level degrees of freedom. For a serial manipulator with n joints, the control input is an n-dimensional vector specifying target joint positions (angles), velocities, or torques at each control timestep. This contrasts with task-space control, which specifies targets in the end-effector's Cartesian frame.

The three main joint-space control modes are position control (commanding target joint angles, with the low-level controller tracking these targets via PID), velocity control (commanding target joint velocities), and torque control (commanding joint torques directly, enabling the most flexible but most demanding control). Position control is dominant in learned manipulation because it is the most stable and the most natural output for behavioral cloning — the policy predicts the joint configuration the robot should be in.

For learned policies, joint-space actions have the advantage of being unambiguous (no IK solver needed, no redundancy resolution), directly executable (the robot's servo-level controller tracks the commanded joint angles), and smooth (the policy can be trained to output smooth joint trajectories). The disadvantage is embodiment specificity — a policy trained with Franka Panda joint-space actions cannot be deployed on a UR5 without retraining.

For teams collecting manipulation data, joint-space labels are automatically available from the robot's joint encoders — they require no additional computation or instrumentation. The practical concerns are encoder accuracy (which is excellent on modern robots, typically sub-degree resolution), consistent homing (the robot must start from the same configuration across sessions), and documentation of joint ordering conventions (different robots may number their joints differently).

Claru's data collection protocols include standardized homing procedures, verified encoder calibration, and documented joint conventions. Each dataset ships with both joint-space and task-space action labels, enabling downstream teams to train in either space without re-collection.

Claru records joint-space labels at native control frequency alongside Cartesian labels in every demonstration dataset, with documented homing procedures and joint conventions for each robot platform.

Joint-space control refers to controlling a robot by directly specifying target values for each joint in the kinematic chain. For a 7-DoF robot arm like the Franka Panda, this means commanding 7 joint positions (angles in radians), 7 joint velocities (radians per second), or 7 joint torques (Newton-meters) at each control timestep. The robot's low-level controllers then track these targets using PID control, impedance control, or torque-feedforward control at the hardware level.

Joint-space control contrasts with task-space (Cartesian) control, where the commands specify end-effector position and orientation. Task-space commands are more intuitive — 'move the gripper 5cm forward' — but require an inverse kinematics solver to convert to joint commands. Joint-space commands are less intuitive but provide direct, unambiguous control over the robot's configuration with no IK ambiguity.

In learned manipulation, the choice between joint-space and task-space action representations has significant implications. Joint-space actions are embodiment-specific — the same joint angles produce entirely different end-effector motions on different robots. Task-space actions are more transferable but introduce IK-related complexities. Many high-performance single-embodiment policies (ACT, Diffusion Policy on specific hardware) use joint-space actions because they eliminate IK artifacts and provide the learning algorithm with the most direct control signal.

When a learned policy outputs joint-space actions, the training data must contain joint-position or joint-velocity labels at each timestep. During teleoperation, the robot's joint encoders record the actual joint angles at each control cycle, providing ground-truth joint-space labels. These labels are typically recorded at the robot's native control frequency (100-1000 Hz) and may be downsampled to match the policy's inference frequency (10-50 Hz).

Joint-space policies learn smooth motion directly — the policy predicts joint angle trajectories that the robot follows without any intermediate conversion. This is why ACT (Action Chunking with Transformers) and many Diffusion Policy implementations use joint-space actions: the predicted trajectories are directly executable, and temporal smoothness in joint space maps to smooth Cartesian motions. The downside is that the policy is locked to the specific robot it was trained on.

For datasets that support joint-space training, the key quality requirements are accurate joint encoder readings (which are standard on modern robots), consistent zero-positions across demonstrations (the robot must be homed identically before each session), and documentation of joint limits and control mode. Claru's datasets include joint-space labels at the native control frequency alongside end-effector labels, with documented joint limits and control configurations.

Collecting high-quality joint-space training data requires attention to the robot's control mode during teleoperation. If the robot is controlled in joint-position mode, the recorded joint angles represent the commanded targets. If controlled in Cartesian mode (which is common for teleoperation interfaces), the recorded joint angles are the result of the IK solver's choices, which may include discontinuities at singularities or near joint limits.

Consistency in the robot's initial configuration across demonstrations is important for joint-space policies. If the robot starts in different configurations for different demonstrations of the same task, the joint-space trajectories will differ significantly even when the Cartesian motions are identical. Standard practice is to define a fixed home configuration and return the robot to this position before each demonstration.

Claru ensures consistent homing procedures, calibrated joint encoders, and documented control modes across all demonstration sessions. Joint-space labels are recorded at the robot's native control frequency and provided alongside Cartesian labels for maximum flexibility in downstream training.