Probabilistic Movement Primitives (ProMP)

• Probabilistic Movement Primitives are a Bayesian framework for representing and encoding robot trajectories using low-dimensional basis function expansions.
• They utilize Gaussian distributions over weight spaces, enabling efficient conditioning on constraints and robust generalization across demonstrations.
• Applications include skill learning, human-robot interaction, and rehabilitation, with extensions like DeepProMPs addressing high-dimensional and complex tasks.

Probabilistic Movement Primitives (ProMPs) are a Bayesian framework for representing, learning, and adapting robot motor skills from demonstrations. They model trajectories as distributions over low-dimensional weight spaces defined by temporal basis functions, enabling the encoding of variability, efficient conditioning on constraints, sequential and online adaptation, and robust generalization. ProMPs have established themselves as a foundational paradigm for movement representation and skill modulation in both robotics and human–robot interaction.

1. Mathematical Definition and Core Representation

A Probabilistic Movement Primitive posits that every trajectory $y(t)\in\mathbb{R}^D$ (joint angles, end-effector poses, etc.) can be described by a time-varying basis-function expansion: $$y(t) = \Phi(t)^\top\,w + \epsilon,\qquad \epsilon\sim\mathcal{N}(0,\,\Sigma_y)$$ Here, $\Phi(t)\in\mathbb{R}^{M\times D}$ denotes a (block-)design matrix comprising $M$ time-varying basis functions (commonly Gaussian RBFs or polynomials) for each degree of freedom, and $w\in\mathbb{R}^{MD}$ is a vector of weights parameterizing the trajectory. Across demonstrations, ProMPs impose a Gaussian distribution on the weights: $w \sim \mathcal{N}(\mu_w,\,\Sigma_w)$ The marginal distribution over trajectories at time $t$ becomes

$$p(y(t)) = \mathcal{N}\left(y(t)\mid\Phi(t)^\top\mu_w,\,\Phi(t)^\top\Sigma_w\Phi(t)+\Sigma_y\right)$$

The entire trajectory $y_{1:T}$ is then modeled as a multivariate normal random process, parameterized by the demonstration-derived mean $\mu_w$ and covariance $\Sigma_w$, which jointly encode nominal motion and variability (Duan et al., 2018, Gutierrez et al., 17 Dec 2025).

2. Parameter Estimation and Learning Algorithms

Batch and Incremental Learning

Given $N$ demonstrations, the standard approach is to extract weight vectors $\{w_n\}$ by ridge regression or EM fitting: $$w_n = \arg\min_w \sum_{t=1}^T \lVert y^{(n)}_t - \Phi_t w \rVert_{\Sigma_y^{-1}}^2 + \lambda\lVert w \rVert^2$$ The sample mean and covariance are estimated as

$$\hat{\mu}_w = \frac{1}{N}\sum_n w_n,\qquad \hat{\Sigma}_w = \frac{1}{N}\sum_n (w_n - \hat{\mu}_w)(w_n - \hat{\mu}_w)^\top$$

Alternatively, EM or MAP-EM can be used, especially under priors (e.g. Normal–Inverse–Wishart) for robustness in low-data regimes (Gomez-Gonzalez et al., 2018).

Incremental algorithms update sufficient statistics with each new demonstration, applying a forgetting factor for continual adaptation: $$u \leftarrow (1-\alpha)u + \alpha u',\qquad \alpha = (N+1)^{-\beta},\,\, 0.5<\beta\leq 1$$ This allows online shaping of $\mu_w,\,\Sigma_w$—essential for interactive learning and adaptation in settings such as human-robot cooperation (Schäle et al., 2021).

High-Dimensional, Riemannian, and Deep Extensions

Orientation ProMPs generalize the weight-Gaussian model from Euclidean space to the sphere $S^3$ for quaternion-valued trajectories, using multilinear geodesic regression on manifolds (Rozo et al., 2021). Neural and deep generative variants, such as DeepProMPs, embed ProMPs in latent-variable models with Bayesian aggregation for high-dimensional or context-rich tasks (Przystupa et al., 2023). Bayesian encoders permit simultaneous conditioning on context variables, via-points, and support for multimodal and sequence-segmented data (Gao, 2024).

3. Modulation, Conditioning, and Adaptation

Conditioning on Constraints (Via-Points, Goals)

A defining feature of ProMPs is analytic conditioning of the weight distribution to satisfy linear constraints: $$\text{Given:}\quad y^* = \Phi^* w + \epsilon,\quad \epsilon\sim \mathcal{N}(0,\Sigma^*)$$ the posterior over $w$ is

$$\Sigma_w^* = \left( \Sigma_w^{-1} + (\Phi^*)^\top{\Sigma^*}^{-1}\Phi^* \right)^{-1}$$

$$\mu_w^* = \Sigma_w^* \left( \Sigma_w^{-1}\mu_w + (\Phi^*)^\top{\Sigma^*}^{-1}y^* \right)$$

Resultant trajectories pass through user-specified via-points or goals while respecting the demonstration-derived variability (Duan et al., 2018, Gutierrez et al., 17 Dec 2025).

Shape and Semantics of Constraints

The conditioning formalism extends to:

• Temporal constraints (phase-dependent via-points) (Lippi et al., 2022).
• Contextual or task descriptors (transfer between tasks) (Stark et al., 2019).
• Arbitrary probabilistic constraints by Kullback–Leibler minimization under chance constraints—e.g. joint/position limits, hyperplane boundaries, obstacle repellers, mutual avoidance, smoothness regularization (Frank et al., 2021).

Blending

Conditioned or context-specific primitives can be product-blended, with time-varying activation functions $a_i(t)$, yielding closed-form expressions for the resulting mean and covariance, both in Euclidean and Riemannian settings (Duan et al., 2018, Rozo et al., 2021).

4. Applications and Specialized Frameworks

ProMPs are used for:

• Skill learning and adaptation: End-effector insertion under tight tolerances, with variance-gated residual reinforcement learning for high-precision tasks (Carvalho et al., 2022).
• Human–robot interaction: Anticipatory motion generation in collaborative settings (IProMPs), with dynamic Kalman updates from human motion and joint phase estimation to maximize robot responsiveness (Duan et al., 2018).
• Rehabilitation robotics: Safe, individualized movement planning under interactive human feedback and variable impedance control (Chen et al., 2023).
• Experience transfer: Knowledge-guided initialization of trajectory distributions for efficient skill acquisition in new tasks (Stark et al., 2019).
• Imitation learning and recognition: Behavioral modeling in driver simulation, human motion analysis under perturbations, and recognition/reproduction under phase uncertainty (Löckel et al., 2020, Xue et al., 2021, Lippi et al., 2022).
• Active learning: Mahalanobis-distance–guided sampling of task contexts for data-efficient construction of primitive libraries (Conkey et al., 2019).
• Hierarchical and segmented motion libraries: Unsupervised segmentation via deep autoencoder/RNN and spectral clustering to automatically decompose motion into reusable primitives (Gao, 2024).

5. Variants and Unified Models

ProDMPs and Dynamics Integration

ProMPs are unified with dynamic movement primitives (DMPs) in the ProDMP framework, replacing online ODE integration with closed-form basis functions derived from the underling DMP dynamics. This yields trajectory distributions that admit both the statistical properties of ProMPs and the goal-attractor, smooth convergence property of DMPs, suitable for end-to-end deep learning and online replanning (Li et al., 2022).

Deep Probabilistic Movement Primitives

DeepProMPs replace the linear basis expansion with a neural decoder and combine contextual and via-point conditioning through Bayesian aggregation in latent space. This generalization permits:

• Higher expressivity and multimodality.
• Scalability to complex, high-dimensional, and vision-conditioned tasks.
• Full support for classical ProMP operations: temporal modulation, via-point and context conditioning, and blending (Przystupa et al., 2023).

6. Empirical Impact and Responsiveness

Empirical studies consistently show that ProMP-based frameworks outperform deterministic or non-adaptive methods in terms of:

• Task generalization with few demonstrations (Conkey et al., 2019, Gomez-Gonzalez et al., 2018).
• Responsiveness and accuracy in collaborative tasks, with endpoint errors reduced by up to 50% under dynamic observation (Duan et al., 2018).
• Smoothness and passivity in contact-rich RL, with significantly reduced jerk and safer energy profiles (Huang et al., 17 Nov 2025).
• Behavioral fidelity in human-robot interaction and imitation (Löckel et al., 2020, Krishna et al., 2022).

7. Limitations, Challenges, and Future Directions

ProMPs exhibit several known limitations:

• The Gaussian formulation limits representation of multimodal or highly nonlinear distributions; future work points toward mixture models, kernel methods, or deep generative extensions (Gomez-Gonzalez et al., 2018, Przystupa et al., 2023).
• Choice and number of basis functions remains a crucial modeling hyperparameter affecting expressivity and stability (Gutierrez et al., 17 Dec 2025).
• Precise time alignment and phase inference are critical; methods leveraging explicit phase distributions or unsupervised segmentation have yielded improvements (Lippi et al., 2022, Gao, 2024).
• High-dimensional and multi-modal human-robot tasks motivate the use of Riemannian methods for orientations and end-to-end neural representations to handle more complex contexts (Rozo et al., 2021, Przystupa et al., 2023).

ProMPs remain an active and evolving foundation for representing, learning, adapting, and controlling movement in robotics and interactive systems. Their probabilistic structure enables closed-form operations critical for real-time and data-efficient adaptation, with ongoing advances focusing on broader expressivity, context integration, and scalability.