Papers
Topics
Authors
Recent
Search
2000 character limit reached

Self-Direction: Definition and Applications

Updated 28 October 2026
  • Self-direction is the capacity of a system, organism, or learner to determine and regulate its own trajectory under constraints, involving autonomy in goal selection and adaptation, such as in autonomous physical systems and self-directed learning.
  • Self-direction can manifest at various levels, including self-direction manifestations in quantum mechanics, hydrodynamics, and self-similar measures, demonstrating both agentic and non-agentic forms.Practical applications of self-direction are found in machine learning, where systems can select tasks and adapt learning strategies; in learning frameworks like Self-Directed Growth, suggesting learners as the primary agents of their developmental trajectory; and in assistive technologies, preserving the meaningful control of individuals with disabilities.
  • Self-direction is a vital concept with practical applications in various domains, including autonomous physical systems, machine learning, educational theory, and assistive technologies. By preserving meaningful agency and adaptation, it significantly influences goal selection, optimization, and overall system performance.

Self-direction is the capacity of a system, organism, learner, agent, community, or dynamical process to determine and regulate its own trajectory under constraints. Depending on the domain, it denotes a monotonically decreasing dynamical observable, autonomous goal and strategy selection, internally generated directional behavior, collective self-organization, learner agency, digital self-determination, or emergent representational structure. Across these uses, self-direction is distinguished from simple automation: it involves some combination of evaluation, orientation, adaptation, feedback, and the preservation of meaningful agency.

1. Conceptual scope and distinguishing properties

Self-direction has no single domain-independent formal definition. In autonomous physical systems, it denotes the ability to evaluate circumstances, execute and achieve goals, interpret instructions, adapt to imperfect information, and respect laws, conventions, and human preferences (Abbeel et al., 2016). In self-directed machine learning, it extends beyond parameter optimization to the selection of tasks, data, models, optimization strategies, and evaluation metrics (Zhu et al., 2022). In educational theory, Self-Directed Growth denotes the sustained capacity to initiate, assess, and revise learning pathways across tasks and contexts (Mao, 29 Apr 2025).

Several recurrent properties distinguish self-direction from narrower forms of automation:

  • Internal orientation: the system possesses, or constructs, a variable, representation, objective, or process that determines a direction of change.
  • Closed-loop adaptation: observations or outcomes modify subsequent action, strategy, or state.
  • Goal or trajectory regulation: the system selects, interprets, or revises goals, subgoals, or developmental pathways.
  • Constraint sensitivity: self-direction operates within physical, social, legal, computational, or architectural constraints.
  • Agency or causal efficacy: the system or person can influence subsequent states rather than merely undergo externally imposed transitions.
  • Emergence: collective or representational self-direction may arise from local rules, optimization pressures, or structural constraints without centralized control.

The term is therefore used at several analytical levels. At the physical level, a reciprocal swimmer can reverse its motion through inertial fluid dynamics without changing its stroke geometry (Dombrowski et al., 2018). At the computational level, a model can infer whether a reasoning step should be reinforced from its own belief change while using a teacher only for token allocation (Li et al., 28 Sep 2026). At the social and political level, digital self-determination concerns the ability of individuals and communities to influence data practices throughout the data life cycle (Verhulst, 2022).

Self-direction should not be conflated with complete independence. Autonomous physical systems may receive demonstrations, instructions, coaching, or cloud computation while remaining self-directed at execution time (Abbeel et al., 2016). Similarly, assistive technology can support self-determination when it preserves a person’s meaningful control rather than eliminating support relationships (Laronze et al., 20 Nov 2025). A plausible implication is that self-direction is better characterized by control over relevant decisions and adaptations than by the absence of external assistance.

2. Dynamical and physical forms

Quantum Lyapunov ordering

In standard quantum mechanics, self-direction is represented by a bounded self-adjoint Lyapunov operator whose expectation value decreases under forward evolution. For a Hamiltonian HH, the unitary group is

U(t)=e−iHt.U(t)=e^{-iHt}.

For a suitable operator MM, the expectation value satisfies

(ψ(t2),Mψ(t2))≤(ψ(t1),Mψ(t1)),t2≥t1≥0.(\psi(t_2),M\psi(t_2))\leq(\psi(t_1),M\psi(t_1)), \qquad t_2\geq t_1\geq0.

The operator is not a canonical time observable and does not satisfy [H,M]=i[H,M]=i. Its role is to order states along a forward trajectory without modifying unitary Schrödinger dynamics (Strauss et al., 2011).

For a Hamiltonian with absolutely continuous spectrum R+\mathbb R^+ and uniform multiplicity, the operator is constructed as the compressed Hardy projection

M=(PR+P+PR+)∣L2(R+;K).M=\left(P_{\mathbb R^+}P_+P_{\mathbb R^+}\right)\big|_{L^2(\mathbb R^+;\mathcal K)}.

It is positive, contractive, injective, bounded, self-adjoint, and has dense range. Its spectrum is purely absolutely continuous and equals [0,1][0,1]. The generalized eigenstates are Mellin-type energy functions, rather than ordinary square-integrable vectors. The expectation value tends to zero as t→∞t\to\infty, providing an intrinsic ordering parameter for the evolving state.

This construction illustrates a non-agentive form of self-direction: the system does not select actions, but its internal dynamics provide a scalar that distinguishes forward temporal position. The result is a structural directionality rather than an assertion that microscopic time-reversal invariance is violated.

Inertial hydrodynamic direction selection

A reciprocal two-sphere swimmer, or spherobot, demonstrates physical self-direction through steady streaming. The swimmer consists of unequal spheres oscillating in antiphase with separation

d(t)=d0+Asin⁡(ωt).d(t)=d_0+A\sin(\omega t).

At zero Reynolds number, reciprocal actuation produces no locomotion by Purcell’s scallop theorem. At finite inertia, nonlinear oscillatory flows generate a nonzero time-averaged steady-streaming field. The asymmetric momentum flux produces propulsion even though the stroke remains geometrically reciprocal (Dombrowski et al., 2018).

The relevant control parameter is

U(t)=e−iHt.U(t)=e^{-iHt}.0

where U(t)=e−iHt.U(t)=e^{-iHt}.1 is the small-sphere amplitude, U(t)=e−iHt.U(t)=e^{-iHt}.2 its radius, U(t)=e−iHt.U(t)=e^{-iHt}.3 the oscillation frequency, and U(t)=e−iHt.U(t)=e^{-iHt}.4 the kinematic viscosity. The simulations identify three regimes:

  1. small-sphere-leading motion at low nonzero U(t)=e−iHt.U(t)=e^{-iHt}.5;
  2. nearly zero mean motion around U(t)=e−iHt.U(t)=e^{-iHt}.6;
  3. large-sphere-leading motion above the transition.

The transition corresponds to reversal of the steady-streaming topology. At lower U(t)=e−iHt.U(t)=e^{-iHt}.7, flow is directed inward along the axis and outward transversely; at higher U(t)=e−iHt.U(t)=e^{-iHt}.8, it is expelled axially and drawn inward transversely. The geometry and reciprocal stroke remain unchanged, while the fluid response selects the direction.

This is a form of hydrodynamic self-direction rather than cognitive autonomy. Geometric asymmetry supplies directional bias, and inertia determines which flow topology dominates.

Active-filament polarity and reversal

In active-filament systems, self-direction is encoded by a polarity variable U(t)=e−iHt.U(t)=e^{-iHt}.9 that determines propulsion along the local tangent. Without reversals, persistent polarity supports two principal structures: spirals in flexible, strongly active filaments and aligned clusters in stiff-filament assemblies (Abbaspour et al., 2021).

The model varies bending stiffness MM0, self-propulsion force MM1, and reversal rate MM2. Reversals occur as a Poisson process, with mean reversal time MM3 and mean free-propagation distance

MM4

Persistent polarity allows flexible filaments to wind around themselves after head–body contact. In stiff assemblies, collisions promote alignment and coherent cluster motion. Reversals disrupt both mechanisms: they drive spirals outward and unwind them, while reversed filaments become incompatible with cluster motion and tend to leave clusters.

Increasing MM5 therefore produces a progression from spirals or large clusters to smaller clusters and mixed states, and ultimately to an isotropic active melt. Here self-direction is not only the ability to move but the persistence of a directional memory that supports structure formation.

3. Collective, geometric, and representational self-direction

Collective self-organization in lattice traffic

The stochastic MM6-dimensional BML model defines particle direction through particle type. A type-MM7 particle attempts to move in the positive MM8-th coordinate direction, while particles can change type with probability MM9 (Kozlov et al., 2023). Movement is local and constrained by exclusion and sequential conflict resolution.

The system is said to reach free movement when every particle continues to move indefinitely. In the Buslaev-net formulation, self-organization corresponds to a velocity single-point spectrum

(ψ(t2),Mψ(t2))≤(ψ(t1),Mψ(t1)),t2≥t1≥0.(\psi(t_2),M\psi(t_2))\leq(\psi(t_1),M\psi(t_1)), \qquad t_2\geq t_1\geq0.0

The principal sufficient condition is

(ψ(t2),Mψ(t2))≤(ψ(t1),Mψ(t1)),t2≥t1≥0.(\psi(t_2),M\psi(t_2))\leq(\psi(t_1),M\psi(t_1)), \qquad t_2\geq t_1\geq0.1

where (ψ(t2),Mψ(t2))≤(ψ(t1),Mψ(t1)),t2≥t1≥0.(\psi(t_2),M\psi(t_2))\leq(\psi(t_1),M\psi(t_1)), \qquad t_2\geq t_1\geq0.2 is the particle number and (ψ(t2),Mψ(t2))≤(ψ(t1),Mψ(t1)),t2≥t1≥0.(\psi(t_2),M\psi(t_2))\leq(\psi(t_1),M\psi(t_1)), \qquad t_2\geq t_1\geq0.3 are the lattice dimensions. The proof introduces diagonal phase classes based on

(ψ(t2),Mψ(t2))≤(ψ(t1),Mψ(t1)),t2≥t1≥0.(\psi(t_2),M\psi(t_2))\leq(\psi(t_1),M\psi(t_1)), \qquad t_2\geq t_1\geq0.4

Movement in any coordinate direction advances the diagonal index by one. Delays shift an associated phase, while persistent phase gaps cannot be eliminated under sufficiently low loading. The resulting global organization does not require centralized route assignment. It emerges from periodic geometry, ordered exclusion, phase invariants, and particle scarcity.

The theorem gives a sufficient rather than necessary condition for the general model. When the lattice dimensions have greatest common divisor one, the condition may provide no nontrivial guarantee, without implying that self-organization is impossible.

Directional structure in self-similar geometry

For self-similar sets, self-direction concerns the recurrence of normal orientations under repeated similarities. A self-similar attractor generated by maps

(ψ(t2),Mψ(t2))≤(ψ(t1),Mψ(t1)),t2≥t1≥0.(\psi(t_2),M\psi(t_2))\leq(\psi(t_1),M\psi(t_1)), \qquad t_2\geq t_1\geq0.5

has spatial and directional structure determined by contraction ratios and orthogonal components (ψ(t2),Mψ(t2))≤(ψ(t1),Mψ(t1)),t2≥t1≥0.(\psi(t_2),M\psi(t_2))\leq(\psi(t_1),M\psi(t_1)), \qquad t_2\geq t_1\geq0.6 (Bohl et al., 2011). The compact group

(ψ(t2),Mψ(t2))≤(ψ(t1),Mψ(t1)),t2≥t1≥0.(\psi(t_2),M\psi(t_2))\leq(\psi(t_1),M\psi(t_1)), \qquad t_2\geq t_1\geq0.7

records the rotations and reflections generated by the iterated function system.

Logarithmically averaged curvature-direction measures separate into a spatial Hausdorff factor and a directional fibre measure:

(ψ(t2),Mψ(t2))≤(ψ(t1),Mψ(t1)),t2≥t1≥0.(\psi(t_2),M\psi(t_2))\leq(\psi(t_1),M\psi(t_1)), \qquad t_2\geq t_1\geq0.8

Under the stated regularity, integrability, and separation assumptions, the directional distribution is almost surely independent of spatial location after orientation is accounted for. Ergodicity of the extended symbolic shift explains why the same asymptotic directional statistics recur at almost every point.

This result does not assert that every finite-scale approximation has the same directional distribution. It describes an asymptotic, logarithmically averaged pattern in which spatial mass and normal-direction statistics decouple.

Directional consistency of self-similar measures

For planar homogeneous self-similar measures with strong separation and dense rotations, self-direction appears as consistent behavior under every one-dimensional projection (Rapaport, 2018). If the measure has dimension greater than one and the contraction parameter lies outside a Hausdorff-dimension-zero exceptional set, then there exists (ψ(t2),Mψ(t2))≤(ψ(t1),Mψ(t1)),t2≥t1≥0.(\psi(t_2),M\psi(t_2))\leq(\psi(t_1),M\psi(t_1)), \qquad t_2\geq t_1\geq0.9 such that every projection [H,M]=i[H,M]=i0 belongs to [H,M]=i[H,M]=i1.

The conclusions include:

  • full projection dimension in every direction;
  • absolute continuity and [H,M]=i[H,M]=i2 regularity of every projection;
  • weak continuity of [H,M]=i[H,M]=i3 in [H,M]=i[H,M]=i4;
  • exact-dimensional conditional measures of dimension [H,M]=i[H,M]=i5;
  • equivalence between each projected measure and Lebesgue measure on the projected attractor.

Under the stronger condition [H,M]=i[H,M]=i6, the projections have Sobolev regularity and vary continuously in the norm topology of [H,M]=i[H,M]=i7. Dense rotations propagate estimates obtained in one direction across the entire unit circle.

Here self-direction denotes directional coherence of a measure: the projection and slicing structure remain dimensionally compatible in every direction, rather than the measure selecting a direction through agency.

4. Learning, autonomy, and artificial systems

Autonomous physical systems

An Autonomous Physical System must independently evaluate circumstances, interpret goals, select and execute actions, monitor consequences, adapt to errors, and comply with rules, laws, conventions, and human instructions (Abbeel et al., 2016). The relevant information may originate from sensors, demonstrations, direct instruction, expert supervision, simulation, other robots, or cloud resources.

The paper identifies three complementary research paths:

  • Deep reinforcement learning for sequential perception-action problems and experience-based policy improvement;
  • Human-robot interaction for demonstrations, coaching, tacit knowledge, causal explanations, and correction;
  • Cloud robotics for shared data, computation, maps, trajectories, and collective robot learning.

Self-direction is stronger than automation because it requires interpretation and adaptation rather than fixed procedural replay. It is also stronger than mere independence from humans: an agent may use human instruction or cloud computation while remaining autonomous in context-sensitive execution.

The principal unresolved problems include safe exploration, data requirements, goal setting, interpretability, unreliable connectivity, privacy, security, mutual trust, and graceful degradation when external infrastructure is unavailable.

Self-directed machine learning

Self-directed machine learning expands the scope of AutoML by allowing a system to select not only models and hyperparameters but also tasks, data, optimizers, and evaluation metrics (Zhu et al., 2022). The proposed framework contains self-awareness and a self-directed learning process linked by feedback:

[H,M]=i[H,M]=i8

Internal awareness includes machine capacity, current state, task knowledge, task dependencies, and learning strategies. External awareness represents multimodal entities, relations, commonsense knowledge, and domain knowledge.

The framework is formulated as a six-level multi-level optimization problem:

  1. self-awareness construction;
  2. task selection;
  3. data and model selection;
  4. optimizer selection;
  5. weight training;
  6. validation and meta-parameter updating.

The formulation remains a proposal rather than a demonstrated general-purpose implementation. The paper does not specify complete algorithms for metric selection, efficient global optimization, discrete task-DAG search, or stable awareness updating. Its principal contribution is to treat the design of the learning process itself as an object of machine learning.

Self-directed reasoning and credit assignment

Decoupled Credit Self-Distillation separates the direction of local reinforcement from its magnitude (Li et al., 28 Sep 2026). The method argues that terminal reinforcement signals provide reliable trajectory-level credit but insufficient local attribution, whereas teacher supervision can provide dense local signals but may couple direction and magnitude to teacher judgment errors and preference variation.

For a reasoning step [H,M]=i[H,M]=i9, the method estimates direction using the change in the student’s belief margin between a correct answer and competing candidates:

R+\mathbb R^+0

A sufficiently positive or negative margin change determines the local direction R+\mathbb R^+1; otherwise, the trajectory-level advantage supplies the fallback. Magnitude is estimated separately through marginal information gain from hidden representations. Teacher information is restricted to within-step token allocation.

The resulting system is not fully autonomous. It requires the correct answer, a terminal verifier, externally selected thresholds, meaningful step segmentation, and teacher supervision for token allocation. Its strongest characterization is student-derived, verifier-anchored, semi-autonomous direction selection.

Direction-aware representation learning

Several machine-learning systems use self-direction to denote internally learned directional representations rather than autonomous goal selection. VectorPOSE uses crop geometry to generate vectors from crop origins to a reference landmark, encoding position, orientation, scale, and extent without manual spatial labels (Zhang et al., 2022). Its Vector Prediction task is combined with Boundary-Focused Reconstruction, which preserves voxel intensities and anatomical boundaries.

DaCCN for monocular depth estimation learns unequal treatment of horizontal and vertical information through trainable directional sampling and cumulative vertical aggregation (Han et al., 2023). The method uses photometric reconstruction rather than explicit direction labels. Its learned vertical and horizontal scales adapt to the environmental dependency of depth, particularly the connection region between the camera and an object.

A spatiotemporal TDANN demonstrates another representational form: direction-selective maps and pinwheel structures emerge from contrastive learning on naturalistic videos combined with a spatial locality loss (Gu et al., 12 May 2026). The optimization balances discriminative pressure and topographic smoothness. Direction-selective responses arise without direction labels, while intermediate spatial regularization produces coherent maps and a residual axial component.

These cases show that “self-direction” may refer to the emergence of a directional latent variable from data, geometry, or inductive bias. It does not necessarily imply agency, goal selection, or autonomous control.

5. Learner, personal, and digital self-determination

Self-Directed Growth

Self-Directed Growth extends self-directed learning from task completion to a sustained developmental trajectory (Mao, 29 Apr 2025). It positions the learner as the initiator, executor, assessor, and primary agent of learning. The framework combines:

  • Capacity to Aspire, which links immediate goals to broader personal, professional, social, and cultural futures;
  • Complex Thinking, which supports analytical, systemic, critical, innovative, scientific, and problem-solving activity;
  • Summative Self-Assessment, which interprets performance and informs the next learning cycle.

The Aspire to Potentials for Learners model organizes these components across person, task, and Person × Task levels. A typical cycle involves articulating aspirations, identifying goals and gaps, selecting abilities, co-creating a complex task with Generative AI, completing the task, evaluating procedures and outcomes, revising a self-assessment framework, and designing the next learning activity.

The framework distinguishes learner agency from task engagement, self-efficacy, or successful completion. Evidence of growth should include sustained goal selection, strategy revision, feedback use, transfer across contexts, recognition of gaps, and persistence across multiple cycles.

The paper is conceptual and does not provide a validated scoring function or production implementation. Its proposed Aspire to Potential Scoring Rubric includes component relevance, relational logistics, and vectorized progression, but no complete algorithm, weighting formula, or statistical validation is supplied.

Assistive technologies and self-determination

For people with intellectual disability and autism spectrum disorder, self-determination is organized around autonomy, self-regulation, psychological empowerment, and self-realization (Laronze et al., 20 Nov 2025). Self-direction does not require doing everything without support. A person remains self-determined when assistance, advice, prompts, or environmental technologies preserve meaningful choice.

Assistive technologies support self-direction through communication, learning, navigation, self-management, home living, work, education, and social participation. Examples include augmentative and alternative communication systems, mobile task prompts, GPS and augmented-reality navigation, self-monitoring applications, video modeling, smart-home platforms, and digital life-planning tools.

The assistant “It’s my life! I choose it” directly supports preference expression, values clarification, future projection, goal prioritization, identification of support needs, and construction of a life plan. DomAssist combines motion, contact, and electricity-consumption sensors with tablets to support safety, reminders, and social participation in the home.

Reported findings include greater richness of life projects, perceived self-determination, psychological well-being, self-esteem, and reduced worry for the life-planning assistant, and improvements in home skills, self-determination, self-esteem, quality of life, and psychological well-being in a six-month DomAssist comparison. The evidence base remains heterogeneous, and many cited studies involve small samples, case studies, or pre–post designs.

Digital self-determination

Digital self-determination applies self-direction to datafied social life. It is defined as respecting, embedding, and enforcing people’s agency, rights, interests, preferences, and expectations throughout the digital data life cycle (Verhulst, 2022).

The framework addresses three asymmetries:

  1. Data asymmetry: unequal access to data and the inability of communities or public-interest actors to obtain relevant information.
  2. Information asymmetry: unequal knowledge about what data is collected, how it is processed, and what consequences may follow.
  3. Agency asymmetry: unequal ability to negotiate, contest, influence, or refuse data practices.

Consent is treated as insufficient because it is often binary, informationally inadequate, individually oriented, and poorly suited to reuse and collective effects. Ownership is also insufficient because data is relational, derived information may concern multiple people, and property models can obstruct public benefit.

The proposed operationalization has four prongs:

  • Processes: data assemblies, public deliberation, participatory impact assessments, and citizen data commons;
  • People and organizations: data stewards and data intermediaries;
  • Policies: charters, social licenses, and codes of conduct;
  • Products and technologies: user-led design and trusted data spaces.

The framework is context-specific but intended to be enforceable across collection, processing, sharing, analysis, and use. Its emphasis is not merely on formal agreement but on symmetric, participatory, negotiated, and accountable data relationships, particularly for vulnerable and marginalized groups.

6. Limitations, controversies, and comparative interpretation

Self-direction is contested because the same term can describe substantially different phenomena. A quantum Lyapunov observable orders states but does not choose goals. A self-propelled swimmer selects motion through fluid dynamics but has no cognitive representation. A self-directed learner exercises agency, while a self-similar measure exhibits directional consistency without agency. Treating all of these as identical would obscure important distinctions.

A useful analytical separation is:

Form Directional source Role of feedback Typical limitation
Dynamical self-direction Hamiltonian, fluid, or lattice dynamics State evolution and physical interaction No intentional goal selection
Representational self-direction Geometry, video statistics, or spatial regularization Reconstruction or contrastive loss Direction may be latent rather than agentive
Agentic self-direction Goals, beliefs, memory, and action policies Outcome monitoring and adaptation Requires robust safety and alignment
Personal self-direction Preferences, aspirations, values, and self-assessment Reflection and developmental feedback Support may become controlling
Digital self-determination Governance, participation, rights, and institutions Negotiation, contestation, and enforcement Collective effects and power asymmetries

Several recurring controversies follow.

Self-direction versus independence: autonomous physical systems and assistive technologies may depend on humans, caregivers, teachers, cloud services, or external data. Dependence does not negate self-direction if the system or person retains meaningful control over decisions.

Emergence versus agency: collective traffic organization, cortical direction maps, self-similar directional measures, and inertial swimming can exhibit organized direction without intention. The term should therefore be qualified as dynamical, emergent, representational, or agentic.

Self-supervision versus autonomy: learning from automatically generated targets does not necessarily imply autonomous goal formation. VectorPOSE and DaCCN use self-generated geometric or photometric supervision, while DCSD still depends on answers, verifiers, thresholds, and teachers. Self-supervision concerns the source of training signals; self-direction concerns control over orientation, goals, or adaptation.

Monotonicity versus strict ordering: the quantum direction-of-time theorem establishes a nonincreasing expectation value, not necessarily strict decrease at every time. The definition’s one-to-one condition applies to non-recurring trajectories, and domain issues prevent treating the formal commutator derivative as the primary rigorous statement (Strauss et al., 2011).

Collective benefits versus individual control: digital self-determination rejects both unrestricted data extraction and simplistic individual ownership. Data may concern communities, and socially valuable reuse may conflict with individual or group preferences. The unresolved problem is to combine agency, public benefit, collective representation, security, and enforceability (Verhulst, 2022).

Performance versus causal explanation: directional learning systems may improve benchmark results without proving that the learned direction corresponds to the intended causal structure. In DCSD, information gain estimates representational novelty rather than exact local causal contribution. In DSLA, synthetic-scene generalization does not establish legal compliance, human acceptability, or closed-loop driving safety (Karlsson et al., 2020).

Across domains, self-direction is most rigorously established when the relevant directional variable, feedback mechanism, constraints, and evaluation criterion are explicit. The concept is weakest when it is used as a synonym for novelty, automation, or performance improvement. A precise account should therefore state what is directing what, from which information, toward which state or objective, under which constraints, and with what evidence of adaptation or agency.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (16)

Topic to Video (Beta)

No one has generated a video about this topic yet.

Whiteboard

No one has generated a whiteboard explanation for this topic yet.

Follow Topic

Get notified by email when new papers are published related to Self-Direction.