Joint Angles

The Degrees of Freedom (DoF) of a robotic joint define its independent directions of motion. A revolute (rotary) joint provides one DoF, measured as a single angle. A robot's total DoF is the sum of its joint DoFs, determining the dimensionality of its configuration space. For example, a typical 6-DoF industrial arm can position its end-effector in 3D space with arbitrary orientation, while a 7-DoF arm offers redundancy, allowing multiple joint configurations for the same end-effector pose.

Forward Kinematics is the function that maps a set of joint angles to the resulting end-effector pose (position and orientation) in Cartesian space. It is computed using a kinematic chain of homogeneous transformation matrices derived from the robot's Denavit-Hartenberg (DH) parameters. This deterministic calculation is foundational for simulation, motion planning, and verifying that a commanded joint state achieves the desired tool position. For a 6-DoF arm, forward kinematics solves: Pose = f(θ₁, θ₂, θ₃, θ₄, θ₅, θ₆).

Physical joint limits are hard constraints on the minimum and maximum angle for each joint, dictated by mechanical stops and wiring. Exceeding them can cause hardware damage. Kinematic singularities are configurations where the robot loses one or more degrees of freedom in Cartesian space, making certain end-effector motions impossible. Common types include:

• Wrist Singularity: When the axes of two wrist joints become aligned.
• Elbow Singularity: When the arm is fully extended or retracted.
• Shoulder Singularity: When the wrist center lies on the axis of the first joint. Motion planners must avoid these regions to maintain controllability.

Robot control and planning occur in two primary spaces:

• Joint Space: The space defined directly by the robot's joint angles (θ₁, θ₂,... θₙ). Motion in this space is straightforward but does not directly describe the end-effector's path.
• Task Space (Cartesian Space): The space defined by the end-effector's position (x, y, z) and orientation (roll, pitch, yaw). Motion here is intuitive for task specification but requires inverse kinematics to translate to joint commands. Trajectory interpolation can be performed in either space. Joint-space interpolation is computationally simple but may cause unpredictable end-effector paths. Task-space interpolation produces straight-line tool paths but requires continuous IK solving.

When representing joint angles for machine learning models, key considerations include:

• Normalization: Angles are typically normalized to a range like [-1, 1] or [0, 1] based on known joint limits to stabilize training.
• Periodicity: For revolute joints, angles like 359° and -1° are physically equivalent. This can be addressed by representing the angle as (sin(θ), cos(θ)) to provide a continuous, periodic signal to the network.
• Temporal Sequencing: Actions are often represented as trajectories—sequences of joint angle vectors over time—which are tokenized for transformer-based action decoders. Velocity and acceleration constraints are implicitly or explicitly enforced to ensure smooth, executable motions.

Joint angles are the output of various high-level control strategies:

• Inverse Kinematics (IK): Solves for joint angles that achieve a desired end-effector pose. Solutions may be multiple (redundancy) or nonexistent (out of workspace).
• Impedance Control: Commands joint torques to achieve a desired dynamic relationship (stiffness, damping) between the end-effector and its environment, often wrapping an inner position loop that uses joint angle feedback.
• Operational Space Control: Directly computes joint torques to achieve commanded end-effector forces and motions, dynamically transforming task-space objectives into joint-space commands. These paradigms demonstrate that joint angles are the final, low-level realization of high-level task intent.

An Inverse Kinematics (IK) Solver is the computational algorithm that calculates the required joint angles for a robotic manipulator to achieve a desired end-effector pose. It solves the inverse of the forward kinematics problem.

• Core Function: Given a target position and orientation for the robot's hand or tool, the IK solver computes the set of joint angles that will place it there.
• Challenges: For redundant manipulators (with more than 6 degrees of freedom), multiple joint angle solutions may exist, requiring optimization for smoothness or obstacle avoidance.
• Example: To pick up a cup, a high-level planner specifies the cup's location. An IK solver translates that into the specific shoulder, elbow, and wrist angles needed for the arm to reach it.

End-Effector Pose is the complete specification of a robot's tool or gripper in Cartesian space, comprising its position (X, Y, Z coordinates) and orientation (often as a quaternion or rotation matrix). It is the primary output of forward kinematics from joint angles.

• Relationship to Joint Angles: The pose is a function of the kinematic chain and its current joint angles. Controlling the robot often involves specifying a desired pose trajectory, which an IK solver converts back into joint angle commands.
• Representation: A 6D vector (3 for position, 3 for orientation) or a 4x4 transformation matrix. In Vision-Language-Action (VLA) models, language instructions like "pick up the blue block" are ultimately grounded as a target end-effector pose.

Cartesian Control is a motion control paradigm where commands are issued directly in the task space (end-effector pose), rather than in joint space. The control system internally computes the necessary joint angle trajectories using real-time Inverse Kinematics.

• Advantage: Intuitive for task specification. A programmer can command the end-effector to move in a straight line or follow a contour, and the controller handles the complex joint coordination.
• Implementation: Often involves a Jacobian matrix, which maps joint velocities to end-effector linear and angular velocities. Resolved-rate motion control is a common Cartesian control technique.
• Use Case: In imitation learning, a human demonstrates a task by moving the robot's end-effector in Cartesian space, recording the pose trajectory for later policy training.

Impedance Control is an advanced strategy that regulates the dynamic relationship between force and motion at the end-effector. Instead of dictating a rigid pose, it defines a virtual spring-damper system, making the robot compliant.

• Core Idea: The controller adjusts the joint torques to achieve a desired mechanical impedance (stiffness, damping, inertia). If the end-effector encounters an unexpected force, it will yield according to the programmed impedance.
• Contrast with Position Control: Pure joint angle or Cartesian position control can lead to high forces on contact. Impedance control is essential for dexterous manipulation and safe human-robot interaction.
• Application: Inserting a peg into a hole, polishing a surface, or physically assisting a human, where precise force interaction is as important as final joint angles.

A Motion Primitive is a parameterized, fundamental movement pattern that serves as a building block for complex robotic behavior. It abstracts a sequence of joint angle trajectories into a reusable skill.

• Abstraction Level: Encodes a short-horizon action like 'reach', 'wipe', or 'screw'. A high-level planner sequences primitives; each primitive generates the low-level joint angle commands.
• Parameterization: A 'reach' primitive might be parameterized by a target end-effector pose. The primitive library contains the know-how to execute the movement smoothly and reliably.
• Role in VLA Models: In action tokenization, a token might represent the initiation of a specific motion primitive. Skill primitives are a closely related concept, often implying a higher-level, goal-directed unit.

A Visuomotor Control Policy is a neural network that maps raw or processed visual observations directly to joint angle commands or motor torques, closing the perception-action loop.

• End-to-End Learning: Bypasses traditional pipelines of object detection, pose estimation, and IK planning. The policy learns the complex mapping from pixels to joint angles through imitation or reinforcement learning.
• Architecture: Often a convolutional neural network (CNN) for vision processing, fused with proprioceptive data (current joint angles), feeding into a multi-layer perceptron that outputs actions.
• Challenge & Promise: While data-hungry and sometimes less interpretable, such policies can learn robust, reactive behaviors for dynamic tasks like grasping from clutter or dexterous manipulation.