Dynamics-Aware Context (DAC)

• DAC is a modeling paradigm that represents changing conditions as explicit contexts, addressing non-stationarity in observations, actions, or features.
• It underpins predictive analytics in manufacturing and model-based reinforcement learning by conditioning models on inferred latent dynamics for improved accuracy.
• DAC enhances adaptive control and decentralized multi-agent coordination by selecting context-specific models, yielding robust performance under dynamic conditions.

Dynamics-Aware Context (DAC) denotes a family of modeling and learning approaches in which context is treated as a representation of the changing conditions that govern how observations, actions, or features should be interpreted. In Manufacturing Operations Management, the idea appears as context-aware analytics built around a semantic context knowledge base whose evolution explains apparent concept drift (Ringsquandl et al., 2014). In reinforcement learning and control, closely related formulations infer latent dynamics regimes from recent transitions or from privileged environment information, then condition dynamics models, policies, or planners on that representation (Lee et al., 2020, Wang et al., 2022, Beukman et al., 2023, Yu et al., 13 Jun 2026). This suggests an umbrella interpretation in which DAC converts non-stationarity, hidden regime variation, or system evolution into an explicit contextual variable that can be modeled, tracked, and used for adaptation.

1. Core problem and formal intuition

A recurring motivation for DAC is that the same observable input can imply different outcomes after the environment or system changes. In the Manufacturing Operations Management formulation, predictive models are written as $M:\mathcal{X}\rightarrow\mathcal{Y}$, trained on labeled source-domain data $D_s$, and concept drift is defined as a mismatch between source and destination conditional distributions,

$$P_s(\mathcal{Y}\mid \mathcal{X}) \neq P_d(\mathcal{Y}\mid \mathcal{X}).$$

The central DAC claim there is: given a sufficiently comprehensive context knowledge base $\mathcal{O}$, the conditional probability distribution $P(\mathcal{Y}\mid \mathcal{O})$ is stationary (Ringsquandl et al., 2014).

In model-based reinforcement learning, the same issue is expressed as a family of dynamics $p_c(s'\mid s,a)$ indexed by a hidden context $c$. Rather than learning a single global predictor $f(s'\mid s,a)$, a context-aware model learns a latent vector from recent transitions and then predicts next state conditioned on that vector (Lee et al., 2020). In contextual-policy formulations, context $c$ selects an MDP $$\mathcal{M}'(c)=\langle\mathcal{S},\mathcal{A},\mathcal{T}^c,\mathcal{R},\gamma\rangle$$, with the important restriction in one formulation that the reward is fixed and only the dynamics vary with context (Beukman et al., 2023). In situational dynamics learning, hidden parameters $D_s$0 alter both transition dynamics and reward structures through a Generalized Hidden Parameter Markov Decision Process $D_s$1 (Murillo-Gonzalez et al., 26 May 2025). In fully decentralized cooperative multi-agent reinforcement learning, DAC formalizes each agent’s locally perceived task as a Contextual MDP in which hidden contexts correspond to distinct joint policies of the other agents (Li et al., 19 Sep 2025).

Across these formulations, the common technical move is to replace a single stationary model with a conditional model: predictions, values, or controls are interpreted relative to a context variable that summarizes the currently active dynamics regime.

2. Semantic context and change tracking in Manufacturing Operations Management

The earliest explicit DAC-style formulation in the supplied literature is the context-aware analytics framework for Manufacturing Operations Management. MOM systems are described as integrating heterogeneous data from automation sources such as sensors and field devices, ERP and enterprise applications such as purchase orders and supplier information, PLM and life-cycle systems, and MES as the manufacturing execution layer. Because these systems evolve continuously, analytics built on top of them face comparability problems over time, concept drift in predictive models, and uncertainty about whether changing predictions are caused by data or by changed underlying context (Ringsquandl et al., 2014).

Context is defined using Dey’s notion as “information that can be used to characterize the situation of an entity.” The corresponding context knowledge base $D_s$2 is an ontology, typically in Description Logic, for example OWL 2 DL / $D_s$3. Its TBox concepts can come from schema-level integration across MOM sources, while ABox assertions reflect dynamic facts from those sources. This separates relatively stable structural knowledge from time-varying factual conditions.

The framework also introduces a formal mapping from historical context models to analytic models,

$D_s$4

where $D_s$5 is the history of context models and $D_s$6 is the set of available analytic models. Model selection is written as

$D_s$7

so the current context determines which analytic model should be used. This yields a context-to-model selection workflow: represent the current system situation as context, compare it to past contexts, and retrieve the model that worked best in the most similar situation.

The paper’s example scenario makes the dynamics-aware aspect concrete. A manufacturer produces two products, $D_s$8 and $D_s$9, both using material $$P_s(\mathcal{Y}\mid \mathcal{X}) \neq P_d(\mathcal{Y}\mid \mathcal{X}).$$0 and welding robot Robo-1. A predictive model is trained to detect whether incoming material will lead to defective products. When Robo-1 is replaced by Robo-2, which is more tolerant and produces fewer defects, the same material quality can lead to different defect labels because the production context changed. The ABox update is written as

$$P_s(\mathcal{Y}\mid \mathcal{X}) \neq P_d(\mathcal{Y}\mid \mathcal{X}).$$1

What appears as drift in the defect concept is therefore explained by a change in context rather than by unexplained statistical variation.

The integration mechanism is explicitly semantic. Automation context can come from OPC UA information models, device descriptions, asset models, and event hierarchies; enterprise context uses ERP and MES data together with ISA-95 / IEC 62264 and B2MML; PLM context uses AutomationML and IEC 62424 CAEX. Local domain standards are semantically lifted to RDF or OWL and unified in a global context ontology. The reported benefits are more homogeneous and comparable data, reuse of models for recurring situations, better detection of incomplete context, and improved predictive accuracy.

3. Latent context inference in model-based reinforcement learning

In model-based reinforcement learning, DAC typically appears as a decomposition between context inference and transition prediction. The Context-aware Dynamics Model, or CaDM, assumes that the environment is a family of MDPs whose transition function changes with a hidden context $$P_s(\mathcal{Y}\mid \mathcal{X}) \neq P_d(\mathcal{Y}\mid \mathcal{X}).$$2, and proposes to infer a latent vector

$$P_s(\mathcal{Y}\mid \mathcal{X}) \neq P_d(\mathcal{Y}\mid \mathcal{X}).$$3

from a short history of recent transitions, then condition a forward dynamics model $$P_s(\mathcal{Y}\mid \mathcal{X}) \neq P_d(\mathcal{Y}\mid \mathcal{X}).$$4 and a backward dynamics model $$P_s(\mathcal{Y}\mid \mathcal{X}) \neq P_d(\mathcal{Y}\mid \mathcal{X}).$$5 on that vector (Lee et al., 2020). The key training objective combines forward and backward prediction,

$$P_s(\mathcal{Y}\mid \mathcal{X}) \neq P_d(\mathcal{Y}\mid \mathcal{X}).$$6

with the same $$P_s(\mathcal{Y}\mid \mathcal{X}) \neq P_d(\mathcal{Y}\mid \mathcal{X}).$$7 used for predictions into nearby future timesteps. The reported rationale is that forcing the context to be useful in both temporal directions encourages it to capture dynamics-specific information rather than superficial features. Empirically, the paper reports that PE-TS + CaDM improves HalfCheetah return in moderate OOD settings from about 2019.6 to 7087.2 (Lee et al., 2020).

ProtoCAD extends the same principle to high-dimensional visual control. It keeps RSSM as the latent transition backbone, adds a prototype-based context module, and constructs a context-aware latent feature

$$P_s(\mathcal{Y}\mid \mathcal{X}) \neq P_d(\mathcal{Y}\mid \mathcal{X}).$$8

where $$P_s(\mathcal{Y}\mid \mathcal{X}) \neq P_d(\mathcal{Y}\mid \mathcal{X}).$$9 is the RSSM latent state, $\mathcal{O}$0 is a projection embedding, and $\mathcal{O}$1 is an aggregated prototype vector (Wang et al., 2022). A central contribution is the temporal crossover SwAV loss $\mathcal{O}$2, which enforces temporal consistency of prototype assignments across time segments of the same latent trajectory. Rather than comparing raw features directly, the method compares prototype assignments across time segments, reflecting the assumption that context does not change within a trajectory segment of a single episode.

ProtoCAD is embedded in a Dreamer-style world-model pipeline. World-model learning updates the RSSM, projector, and prototypes; imagination rolls out latent trajectories and recomputes $\mathcal{O}$3 and $\mathcal{O}$4; actor and critic are then conditioned on $\mathcal{O}$5. The paper reports evaluation on 8 visual control tasks from DM-Control with modified dynamics, with 2M environment steps, action repeat 2, and evaluation every 10K steps on unseen dynamics averaged over 5 episodes. Across all dynamics generalization tasks, ProtoCAD delivers 13.2% better mean performance and 26.7% better median performance than Dreamer (Wang et al., 2022).

Taken together, these methods indicate a characteristic DAC pattern in model-based RL: recent interaction history is converted into a compact latent description of local dynamics, and that description is then used to disambiguate prediction and improve generalization under unseen transition settings.

4. Context-conditioned policies and controllers

A second major line of work conditions the policy or controller itself on a context variable rather than using context only inside the dynamics model.

Paper	Context source	Conditioning mechanism
"Dynamics Generalisation in Reinforcement Learning via Adaptive Context-Aware Policies" (Beukman et al., 2023)	Ground-truth context in experiments	Hypernetwork-generated adapter weights inside policy
"Can Context Bridge the Reality Gap? Sim-to-Real Transfer of Context-Aware Policies" (Iannotta et al., 6 Nov 2025)	Estimated from $\mathcal{O}$6 recent transitions	Policy input $\mathcal{O}$7
"Learning Context-Aware Neural ODE Dynamics for Adaptive Robotic Control" (Yu et al., 13 Jun 2026)	Inferred from state-action histories	Neural ODE dynamics and MPC/MPPI
"Context-aware controller inference for stabilizing dynamical systems from scarce data" (Werner et al., 2022)	Unstable dynamical subspace near a steady state	Feedback restricted to $\mathcal{O}$8

The Decision Adapter architecture conditions behavior on context without concatenating context directly to the state. A base feedforward policy $\mathcal{O}$9 is modified by inserting adapters $P(\mathcal{Y}\mid \mathcal{O})$0, whose parameters are generated from context by hypernetworks,

$P(\mathcal{Y}\mid \mathcal{O})$1

The paper positions this as a generalization of cGate, shows that “middle” adapter locations work best, and reports superior generalization performance together with greater robustness to irrelevant distractor variables than concatenation, cGate, and FLAP in ODE, CartPole, and Mujoco Ant experiments (Beukman et al., 2023).

The sim-to-real transfer study follows a related but explicitly identification-oriented route. A context estimator

$P(\mathcal{Y}\mid \mathcal{O})$2

maps recent transitions to an inferred context representation, and the policy uses $P(\mathcal{Y}\mid \mathcal{O})$3 as input (Iannotta et al., 6 Nov 2025). Three supervision strategies are compared: Ground-Truth supervision, Forward Prediction proxy supervision, and Policy Loss supervision. On the Pendulum benchmark, conditioning on explicit context improves generalization over the agnostic domain-randomization baseline; on the Franka Emika Panda pushing task, context-aware policies outperform the context-agnostic baseline across all reported settings, although the best supervision strategy depends on the task. In the harder real-robot setting with center-of-mass variation, the agnostic baseline has about $P(\mathcal{Y}\mid \mathcal{O})$4 reward and 0.32 success, while FP achieves around $P(\mathcal{Y}\mid \mathcal{O})$5 reward and 0.48 success (Iannotta et al., 6 Nov 2025).

The context-aware Neural ODE controller uses a two-phase training procedure. In Phase 1, privileged environmental information $P(\mathcal{Y}\mid \mathcal{O})$6 is mapped to a latent context $P(\mathcal{Y}\mid \mathcal{O})$7, and a continuous-time dynamics model predicts $P(\mathcal{Y}\mid \mathcal{O})$8. In Phase 2, a separate adaptive module reconstructs the latent context from a window of past state-action pairs,

$P(\mathcal{Y}\mid \mathcal{O})$9

with loss $p_c(s'\mid s,a)$0. The learned dynamics are then used inside MPC solved by MPPI. The reported experiments cover a quadrotor under temporally and spatially varying wind, a Sphero BOLT robot on randomized friction layouts, and a Fanuc manipulator with center-of-mass shifts and disturbances. The paper states that the method adapts to temporally and spatially varying environmental changes across different tasks and reports over 60% error reduction in the extreme planar-pushing case and over 80% under disturbance relative to RMA (Yu et al., 13 Jun 2026).

Context-aware controller inference uses the term differently. Here the relevant context is the unstable dynamical subspace near a steady state. If $p_c(s'\mid s,a)$1 spans the unstable left eigenspace, then controllers of the form

$p_c(s'\mid s,a)$2

are sufficient for local stabilization, so only unstable dynamics need to be learned (Werner et al., 2022). The paper reports stabilizing controllers from orders of magnitude fewer data samples than traditional data-driven control techniques and variants of reinforcement learning, including examples such as 104× fewer samples on heat flow if the basis is known and stabilization of the Brazilian interconnected power system with 2 observations if the basis is known (Werner et al., 2022). This is a narrower but still recognizably DAC-style use of context: only task-relevant dynamics are retained.

5. Situation modeling and decentralized multi-agent DAC

A more explicitly online and non-stationary formulation appears in situational dynamics learning. There, the robot’s “situation” is defined as the joint distribution of the latest state transitions rather than as a latent belief over an unobserved state variable. The method learns the joint distribution

$p_c(s'\mid s,a)$3

using a multivariate Gaussian model over

$p_c(s'\mid s,a)$4

with a Normal-Wishart prior and a multivariate extension of Bayesian Online Changepoint Detection (Murillo-Gonzalez et al., 26 May 2025). A changepoint signals that the stream of transition tuples is no longer explained by the same underlying data-generating process, so a new situation has emerged. The identified distributions are then turned into a symbolic latent state $p_c(s'\mid s,a)$5 using the Gaussian moment-generating-function construction, and the robot’s transition model is conditioned on $p_c(s'\mid s,a)$6.

This situation representation is integrated into a model-based planning pipeline for unstructured terrain navigation. The reported latent factors include friction, ground compliance, surface irregularity, slope-induced dynamics, slippage, and internal motion variation. In simulation, the situationally-aware model reaches 100% success in the first three environments and 86% in the hardest one; in the real world, it is reported as the only method to complete all trials successfully on the mixed-terrain task (Murillo-Gonzalez et al., 26 May 2025). The paper also reports emergent safe behaviors such as backing up, slowing down on steep or uncertain terrain, and choosing curved paths to maintain stability.

In fully decentralized cooperative multi-agent reinforcement learning, DAC is formulated as a context modeling solution to non-stationarity and relative overgeneralization. Each agent observes only the shared state, its own local actions, and the shared rewards, so the local process is written as a family of MDPs indexed by hidden contexts, each context corresponding to a distinct joint policy of the other agents (Li et al., 19 Sep 2025). A VAE-like latent model infers discrete context variables $p_c(s'\mid s,a)$7 from a sliding window of recent local transitions, and a context-conditioned value function $p_c(s'\mid s,a)$8 is updated with context-augmented Bellman targets. For action selection, the method defines the optimistic marginal value

$p_c(s'\mid s,a)$9

which is used to promote cooperative local actions while addressing relative overgeneralization.

The empirical evaluation covers matrix game, modified predator-prey, and SMAC tasks. The reported finding is that DAC outperforms IQL, Hysteretic Q-learning, and I2Q in difficult settings, and remains the only method to succeed robustly in the hardest predator-prey case with $c$0 (Li et al., 19 Sep 2025). In this literature, DAC is therefore both a latent context model and a coordination mechanism: context stabilizes value updates, while optimistic marginalization addresses the action-selection bias induced by unobserved teammate behavior.

6. Benefits, limitations, and terminological scope

Across the cited works, several benefits recur. In MOM, explicit context supports more homogeneous and comparable data, reuse of models for recurring situations, and diagnosis of incomplete context knowledge when drift occurs without observed context change (Ringsquandl et al., 2014). In model-based RL, context-aware dynamics improve prediction accuracy, reduce compounding rollout error, and improve zero-shot generalization across unseen dynamics; ProtoCAD reports 13.2% mean and 26.7% median gains over Dreamer, while CaDM improves robustness to moderate and extreme OOD dynamics (Wang et al., 2022, Lee et al., 2020). In context-conditioned policies, the Decision Adapter is more robust to irrelevant distractor variables, and context-aware sim-to-real policies outperform a context-agnostic domain-randomization baseline across all reported settings (Beukman et al., 2023, Iannotta et al., 6 Nov 2025). In robotics and control, context-aware Neural ODEs adapt to temporally and spatially varying disturbances, situation-aware dynamics learning yields safer adaptive navigation, and unstable-subspace controller inference attains local stabilization from orders of magnitude fewer samples than full-model approaches (Yu et al., 13 Jun 2026, Murillo-Gonzalez et al., 26 May 2025, Werner et al., 2022).

The limitations are equally consistent. The MOM framework explicitly notes that manual construction of semantic context models is elaborate and error-prone, and that drift without context change indicates that the context model may be incomplete (Ringsquandl et al., 2014). The Decision Adapter experiments assume ground-truth context and note sensitivity to badly noisy or badly scaled context (Beukman et al., 2023). In sim-to-real transfer, estimation becomes harder as the hidden dynamics space grows, and no learned estimator surpasses the Oracle baseline that appends the true context directly to the state (Iannotta et al., 6 Nov 2025). The Neural ODE controller is slower at inference than model-free policies and degrades when training and test time steps mismatch (Yu et al., 13 Jun 2026). Context-aware controller inference provides local rather than global guarantees and depends on knowing or estimating the unstable eigenspace basis $c$1 (Werner et al., 2022).

The acronym itself is overloaded. In the supplied literature, “DAC” also appears in dynamic algorithm configuration through “CANDID DAC” (Bordne et al., 2024), in prompt compression as “Dynamic Attention-aware Approach” (Zhao et al., 16 Jul 2025), and in agent systems as “DACmMCMAS” (Costantini, 2014). A plausible implication is that “Dynamics-Aware Context” is best understood not as a single canonical formalism but as a research theme: whenever hidden or changing dynamics make a fixed model brittle, DAC-style methods attempt to externalize those dynamics into an explicit contextual representation and condition analytics, prediction, control, or coordination on that representation.