Dynamic Movement Primitive (DMP)

The foundation of a DMP is a time-independent, stable canonical dynamical system, typically a second-order differential equation like a damped spring. This system provides a phase variable that monotonically drives the movement from start to finish, ensuring temporal consistency and stability regardless of execution speed or external perturbations.

• Key Property: The phase variable replaces explicit time dependence, making the movement robust to temporal scaling and interruptions.
• Mathematical Form: Often modeled as: ẋ = α(β(g - x) - ẋ)/τ, where τ is a temporal scaling factor.

The shape of a specific movement is encoded in a nonlinear forcing function. This function is learned from a demonstration and superimposed onto the canonical system. It is typically represented as a weighted sum of radial basis functions (RBFs) centered along the phase variable.

• Function: f(φ) = (Σ ψ_i(φ) w_i) / Σ ψ_i(φ) * φ, where ψ_i are basis functions and w_i are learned weights.
• Role: This function sculpts the attractor dynamics of the canonical system to reproduce complex trajectory shapes, from simple reaches to rhythmic motions like waving.

DMPs provide explicit, closed-form equations for spatial and temporal scaling, allowing a single learned primitive to adapt to new goals and execution speeds without re-learning.

• Spatial Scaling: The trajectory's amplitude and target (g) can be changed. The forcing function scales appropriately to preserve the movement's shape relative to the new start and goal positions.
• Temporal Scaling: The execution speed is controlled by the parameter τ in the canonical system. Increasing τ slows the movement while preserving its dynamic profile.
• Example: A DMP for a 10cm reach can be instantly adapted to perform a 50cm reach at half speed.

DMPs can react to real-time sensory input through coupling terms. These terms modify the trajectory generation online by adding external forcing to the dynamical system equations, enabling reactive behaviors like obstacle avoidance without corrupting the underlying movement representation.

• Mechanism: A coupling term C is added: ẍ = α(β(g - x) - ẋ)/τ + f(φ) + C(sensor_input).
• Use Case: A repulsive vector field from an obstacle sensor can be used for C, causing the end-effector to deflect smoothly around the obstacle before converging back to the original goal.

The DMP framework has two primary formulations for different motion classes:

• Discrete DMPs: Model point-to-point, finite-duration movements like reaching or placing. They use a critically damped spring canonical system that converges to a single goal state.
• Rhythmic DMPs: Model periodic, continuous movements like walking, drumming, or cranking. They use a limit cycle oscillator (e.g., a modified Hopf oscillator) as the canonical system to generate stable oscillations.

This duality allows the same mathematical framework to encode a wide range of robotic skills.

A critical feature of DMPs is their provable global asymptotic stability. The canonical system's dynamics ensure the trajectory will always converge to the goal g from any initial state, provided the forcing function vanishes at the end of the movement.

• Engineering Benefit: This guarantees that the robot will not diverge to infinity or get stuck in spurious attractors, even if the learned forcing function is imperfect or external perturbations are applied mid-execution.
• Contrast with Pure Playback: Unlike simply replaying a recorded trajectory, the DMP's stability property makes it robust to disturbances, allowing the robot to recover and still reach the goal.

DMPs are a core technique for Learning from Demonstration (LfD). A single human-guided trajectory is recorded and used to learn the parameters of the DMP's nonlinear forcing function. This allows the robot to:

• Generalize the skill to new start or goal positions.
• Reproduce the demonstrated motion profile with temporal and spatial consistency.
• Scale the speed or amplitude of the movement while preserving its shape. This is fundamental for teaching complex manipulation skills like assembly or tool use without explicit programming.

In logistics and warehousing, DMPs enable robust pick-and-place operations. The canonical system provides a reliable baseline motion, while the forcing function encodes the specific lifting and placing dynamics.

• Goal Adaptation: If a target bin's location changes, only the DMP's attractor goal parameter needs updating; the motion profile (acceleration, approach angle) remains intact.
• Obstacle Avoidance: The trajectory can be modulated in real-time by adding coupling terms to the differential equations, allowing the arm to smoothly deviate around unexpected obstacles before converging to the goal.
• Force-Sensitive Placement: For delicate items, the trajectory can be combined with impedance control at the endpoint.

The rhythmic DMP formulation is specifically designed for cyclic movements. Instead of converging to a single goal, it produces stable limit cycles.

• Applications: Polishing, wiping, stirring, cranking, or walking gait cycles for legged robots.
• The system learns the frequency and shape of the oscillation from demonstration.
• Adaptation: The frequency, amplitude, and offset of the oscillation can be modified independently, allowing a stirring motion to adapt to a larger bowl or a wiping motion to cover a bigger surface without re-programming.

DMPs are used to coordinate the complex, contact-rich finger motions required for in-hand manipulation.

• Finger Gaiting: A DMP can generate the trajectory for a finger to break contact, reposition, and re-establish contact during regrasping.
• Object Reorientation: By sequencing DMPs for individual fingers or the wrist, an object can be rolled or pivoted within the hand.
• The stability properties of the DMP's differential equations help ensure smooth force transitions, which is critical to prevent object slip.

DMPs act as a skill representation within hierarchical robot architectures.

• A task and motion planner selects a sequence of high-level actions (e.g., grasp(bolt), insert(bolt, hole)).
• Each action is parameterized by a corresponding DMP (or set of DMPs). The planner outputs the DMP parameters (start, goal, duration).
• This decouples planning from low-level trajectory generation, making the system more modular and responsive. The DMP executes the smooth, goal-directed motion while the planner prepares the next skill.

DMPs are a key tool for bridging the sim-to-real gap in policy learning.

• A policy can be trained in simulation using Reinforcement Learning (RL) or Imitation Learning to output DMP parameters rather than raw joint torques.
• The DMP layer provides a structured action space that enforces smoothness and stability by design, which is more transferable than unstructured torque commands.
• The policy learns which DMP to execute and how to parameterize it, while the DMP itself ensures physically plausible motions upon deployment in the real world.

Trajectory optimization is the process of computing a sequence of robot states and control inputs that minimizes a cost function while satisfying dynamic constraints and task goals. It is a broader mathematical framework than DMPs, which provide a specific, efficient parameterization for such trajectories.

• Key Methods: Include direct collocation and shooting methods.
• Relation to DMPs: A DMP can be seen as a canonical system whose parameters (weights, goal, temporal scaling) are often learned or optimized via trajectory optimization techniques to satisfy task-specific costs.

Model predictive control is an advanced, online control method where a dynamic model of the system is used to predict future behavior and optimize a sequence of control inputs over a receding horizon. It is frequently used for reactive, disturbance-rejecting control.

• Online vs. Offline: MPC solves optimization in real-time, whereas a DMP typically represents a pre-computed or learned motion primitive.
• Synergy: DMPs can provide the reference trajectory for an MPC controller, which then handles real-time deviations and force interactions, combining the benefits of both approaches.

Impedance control is a robot control strategy that regulates the dynamic relationship between force and motion at the end-effector, making the robot behave as a programmable mass-spring-damper system. It defines how the robot reacts to contact.

• Execution Layer: While a DMP generates a desired position or velocity trajectory, impedance control executes it by modulating the robot's mechanical impedance.
• Critical for Dexterity: This combination allows a robot to follow a DMP-generated path while safely and compliantly interacting with uncertain environments.

An inverse kinematics solver is an algorithm that computes the joint angles required for a robotic manipulator to achieve a desired position and orientation of its end-effector in Cartesian space.

• Downstream Processing: A DMP typically generates a smooth trajectory in Cartesian space (or joint space). An IK solver is then used to convert each Cartesian waypoint along the DMP trajectory into a feasible joint configuration for the physical robot.
• Integration: For obstacle avoidance, the IK solver must find solutions that respect the DMP's path while also avoiding joint limits and self-collisions.

In reinforcement learning, a policy network is a neural network that maps observations from an environment directly to actions (or a distribution over actions). It represents a learned control policy.

• Alternative to DMPs: A policy network can learn complex, reactive behaviors from scratch but is often a black box and less predictable.
• Hybrid Approach: DMPs are frequently used as a structured, composable action space within a larger RL policy. The RL agent learns to modulate DMP parameters (goal, shape) based on perception, combining the robustness of learned policies with the stability of DMPs.

Probabilistic Movement Primitives are a Bayesian generalization of DMPs that represent a trajectory as a probability distribution over time. This allows for uncertainty representation, conditioning on via-points, and blending of multiple primitives.

• Key Advancement over DMPs: While a standard DMP represents a single deterministic trajectory, a ProMP represents a distribution of trajectories, enabling:
  • Uncertainty-aware execution and planning.
  • Natural modulation via probabilistic conditioning.
  • Co-activation of multiple primitives.