Steerable Policies: Adaptive Control Interfaces

• Steerable policies are decision rules whose behavior is modulated post-training by an explicit low-dimensional steering interface, enabling flexible control without retraining.
• They are applied across domains such as robotic RL, stochastic control, and language-conditioned systems, employing mechanisms like latent space optimization and covariance steering.
• Practical implementations use methods like contrastive learning, safety-preserving geometric controls, and instruction-conditioned modulation to achieve measurable performance improvements.

Searching arXiv for papers on steerable policies and closely related formulations. Searching for "steerable policies" and adjacent terms in robotics, language policies, and control. Steerable policies are policies whose behavior can be modulated by an explicit steering interface at deployment or inference time rather than being fixed after training. Across recent work, the steering variable may be a latent code, a feedback law, a covariance target, a language instruction, a policy document, a numerical cost vector, or a value-function constraint. The common pattern is a separation between a rich policy class and a structured control surface through which behavior can be selected, adapted, or constrained without retraining the entire system (Wang et al., 15 Jun 2026, Kumagai et al., 29 Jan 2026, Wang et al., 2024, Chen et al., 13 Feb 2026).

1. Conceptual scope

The term is used in several distinct but related senses. In robotic RL, a steerable policy is a frozen generative controller whose latent input is optimized by a separate policy, so that expressive stochastic exploration is converted into deterministic, state-dependent deployment behavior (Wang et al., 15 Jun 2026). In stochastic control, it denotes a state-feedback law that steers the mean and covariance of a state distribution, not merely a nominal trajectory (Kumagai et al., 29 Jan 2026, Bakolas, 2020). In language and vision-language-action systems, it refers to policies conditioned on external descriptors such as objective weights, fine-grained instructions, policy documents, or task-switch commands (Wang et al., 2024, Hu et al., 26 May 2026, Chen et al., 18 Mar 2026). In motion planning and geometric control, steerability appears as curvature-constrained trajectory synthesis, or as low-dimensional action interfaces layered on top of autonomous geometric motion policies with safety guarantees (Fu et al., 2021, Fu et al., 2021, Wu et al., 20 May 2026).

Family	Steering variable	Representative works
Generative robotic control	Latent $z$, deterministic latent actor	SteerGenPO (Wang et al., 15 Jun 2026)
Stochastic control	$v_k, K_k$, disturbance-feedback gains, covariance factors	CS and minimum-variance steering (Kumagai et al., 29 Jan 2026, Bakolas, 2020)
Language and alignment	$(\alpha,w)$, user context, policy text, activation vectors, cost coefficients	CLP (Wang et al., 2024), few-shot alignment (Kobalczyk et al., 2024), CoPE (Chakrabarti et al., 19 Dec 2025), SafeSteer (Ghosh et al., 1 Jun 2025), clarification policies (Berant et al., 3 Dec 2025)
VLAs and multitask robotics	Fine-grained instructions, command abstractions, task-switch prompts	FineVLA (Hu et al., 26 May 2026), steerable VLA control (Chen et al., 13 Feb 2026), ReSteer (Chen et al., 18 Mar 2026)
Safety-critical motion	Motion primitives, task-manifold residual actions	Steerable needles (Fu et al., 2021, Fu et al., 2021), SafePBDS (Wu et al., 20 May 2026)

A recurring misconception is that steerability is equivalent to ordinary task conditioning. Recent robotic evidence distinguishes the two sharply: multitask policies can exceed 90% single-task success while remaining poorly responsive to mid-execution instruction changes, which ReSteer identifies as a distinct failure mode of task steerability (Chen et al., 18 Mar 2026). FineVLA makes the same distinction at the instruction level, separating goal completion from execution-conditioned compliance such as active arm, approach direction, contact region, and final configuration (Hu et al., 26 May 2026).

2. Latent and behavior-space steering

A prominent formulation treats steering as control over a latent interface to a richer policy backbone. In "Steering Generative Reinforcement Learning into Stable Robotic Controller," SteerGenPO first trains a flow-like generative policy with stochastic latent sampling,

$$z \sim p_0(z)=\mathcal N(0,I), \qquad a = D_\theta(s,z),$$

and then freezes the generator and learns a latent actor

$\pi_\phi^Z(z\mid s)=\mathcal N(\mu_\phi(s),\Sigma)$

in the induced latent MDP. At deployment, stochastic latent sampling is replaced by the deterministic latent $z=\mu_\phi(s)$, yielding a deterministic executable policy $a=D_\theta(s,\mu_\phi(s))$ (Wang et al., 15 Jun 2026). The paper explicitly frames this as a concrete instantiation of steerable policies: a stochastic generator defines a rich action manifold, while a low-dimensional latent policy steers that manifold. Empirically, SteerGenPO outperforms PPO and GenPO on all six Isaac Lab tasks, with improvements of up to $88.6\%$ over PPO and up to $16.9\%$ over GenPO, and exhibits more stable path tracking and real-robot Unitree G1 deployment behavior (Wang et al., 15 Jun 2026).

A related but more abstract formulation appears in "Learning Policy Representations for Steerable Behavior Synthesis," which models a policy representation as an occupancy-measure expectation,

$h_\pi=\mathbb E_{d_\pi}[f(s,a)].$

The representation is approximated from unordered sets of state-action samples by a set-based encoder, regularized as a variational latent variable, and shaped with contrastive learning so that latent distances align with value-function differences (Li et al., 29 Jan 2026). The decoder reconstructs a policy $v_k, K_k$0, while per-objective heads predict $v_k, K_k$1. This geometry supports test-time steering by solving a constrained optimization problem directly in latent space,

$v_k, K_k$2

followed by decoding the optimized latent into a policy (Li et al., 29 Jan 2026). This suggests a broader interpretation of steerability: not only steering a policy’s current action, but steering an entire policy instance within a learned manifold of behaviors.

These latent-space approaches share a technical separation between representation and execution. The steering signal is low-dimensional and structured; the realized policy remains high-capacity and nonlinear. That separation is explicit in SteerGenPO’s “exploration versus control” decomposition and implicit in occupancy-based behavior synthesis, where latent optimization replaces retraining (Wang et al., 15 Jun 2026, Li et al., 29 Jan 2026).

3. Distribution steering, motion planning, and geometric safety

In stochastic control, steerability is often distributional. "Square Root-Factorized Covariance Steering" studies memoryless affine state feedback

$v_k, K_k$3

for discrete-time LTV systems with Gaussian noise, where $v_k, K_k$4 steers the mean and $v_k, K_k$5 steers covariance propagation (Kumagai et al., 29 Jan 2026). The paper reformulates chance-constrained covariance steering in terms of Cholesky factors $v_k, K_k$6, propagates them through QR-based square-root dynamics, and solves the resulting non-convexity with sequential convex programming. It proves global optimality in the unconstrained expectation-of-quadratic setting and shows that, with chance constraints, the square-root formulation shares the same local minima as covariance-based formulations (Kumagai et al., 29 Jan 2026). The steering object here is the Gaussian state distribution itself.

"Minimum Variance and Covariance Steering Based on Affine Disturbance Feedback Control Parameterization" uses an affine disturbance-feedback law

$v_k, K_k$7

so that terminal mean and covariance become convex functions of the controller parameters (Bakolas, 2020). Minimum-variance steering reduces to a convex QCQP, while covariance steering becomes an SDP with an LMI constraint on terminal covariance (Bakolas, 2020). The paper also introduces truncated disturbance histories to trade performance against computational cost. In this lineage, a steerable policy is one that shapes first- and second-order statistics under uncertainty, rather than merely tracking a nominal plan.

Safety-critical motion planning provides another interpretation. "Toward Certifiable Motion Planning for Medical Steerable Needles" formulates steerable needle control through motion primitives $v_k, K_k$8 with $v_k, K_k$9, and gives a resolution-complete planner that either finds an exact obstacle-avoiding plan in finite time or certifies non-existence at the chosen resolution (Fu et al., 2021). "Resolution-Optimal Motion Planning for Steerable Needles" strengthens this to resolution-optimality, proving that the returned path cost is within $(\alpha,w)$0 of the globally optimal qualified plan at sufficiently fine resolution (Fu et al., 2021). Here steerability is curvature-constrained reachable-set control with formal guarantees.

"Safe and Steerable Geometric Motion Policies for Robotic Dexterous Manipulation" adds a more explicitly policy-oriented notion. SafePBDS composes task-manifold dynamical systems into configuration-space accelerations, adds a pullback control barrier function construction to convert task-space safety conditions into linear constraints on configuration-space accelerations, and introduces a task-manifold action interface through which a high-level policy injects low-dimensional residual motions (Wu et al., 20 May 2026). Zero input recovers autonomous behavior, while safety is preserved under arbitrary inputs. On dexterous grasping, the method reports a $(\alpha,w)$1 success rate across 20 household objects and 120 trials; with the action interface, it can exclude any one finger during grasping via a one-dimensional action and achieves $(\alpha,w)$2 3-finger grasp success across 3 objects and 36 trials; it also enables fully actuated palm-down in-hand reorientation exceeding $(\alpha,w)$3 of yaw in both directions (Wu et al., 20 May 2026). This is a particularly clear example of a steerable policy as an autonomous controller with a safe residual control port.

4. Language-, preference-, and policy-conditioned steerability

In language modeling, steerability is often realized by conditioning a single policy on external preference or governance descriptors. "Conditional Language Policy: A General Framework for Steerable Multi-Objective Finetuning" defines a conditional policy

$(\alpha,w)$4

parameterized by a KL weight $(\alpha,w)$5 and reward-mixture weights $(\alpha,w)$6, and learns it by multi-task reward-based finetuning over sampled $(\alpha,w)$7 pairs (Wang et al., 2024). The conditioned parameter block is

$(\alpha,w)$8

with $(\alpha,w)$9. The resulting model can trade off objectives continuously at inference time without maintaining separate policies, and the paper reports Pareto-dominance over Rewarded Soups and prompt-only conditioning on summarization benchmarks (Wang et al., 2024).

"Few-shot Steerable Alignment" moves from global objective weights to user-specific preference inference. Each user is represented by a few-shot context set $$z \sim p_0(z)=\mathcal N(0,I), \qquad a = D_\theta(s,z),$$0, encoded into a latent $$z \sim p_0(z)=\mathcal N(0,I), \qquad a = D_\theta(s,z),$$1, which conditions both a reward model and a FiLM-modulated LLM policy:

$$z \sim p_0(z)=\mathcal N(0,I), \qquad a = D_\theta(s,z),$$2

The framework extends Bradley–Terry–Luce preference modeling to heterogeneous latent preferences and supports adaptation to a continuum of behavioral modes from only a few examples of user choices (Kobalczyk et al., 2024). This is steerability through inferred preference codes rather than explicitly chosen reward weights.

Several papers make the steering signal itself a text policy. "CoPE: A Small LLM for Steerable and Scalable Content Labeling" takes a policy document $$z \sim p_0(z)=\mathcal N(0,I), \qquad a = D_\theta(s,z),$$3 and content $$z \sim p_0(z)=\mathcal N(0,I), \qquad a = D_\theta(s,z),$$4 as input and outputs a binary decision relative to that policy, with behavior changing immediately when the policy text changes (Chakrabarti et al., 19 Dec 2025). Its Contradictory Example Training and Binocular Labeling procedures are designed to force policy interpretation rather than policy memorization. "SafeSteer" instead modifies internal activations at inference time using category-specific steering vectors $$z \sim p_0(z)=\mathcal N(0,I), \qquad a = D_\theta(s,z),$$5,

$$z \sim p_0(z)=\mathcal N(0,I), \qquad a = D_\theta(s,z),$$6

to reduce unsafe generations while preserving topic relevance and avoiding blanket refusal (Ghosh et al., 1 Jun 2025). "Learning Steerable Clarification Policies with Collaborative Self-play" conditions a conversational policy on two numerical costs, $$z \sim p_0(z)=\mathcal N(0,I), \qquad a = D_\theta(s,z),$$7 for clarification turns and $$z \sim p_0(z)=\mathcal N(0,I), \qquad a = D_\theta(s,z),$$8 for final-answer word count, and optimizes cost-penalized accuracy via self-play and ReST; the resulting policy changes clarification, enumeration, and verbosity behavior predictably as these costs vary, including for unseen numerical values (Berant et al., 3 Dec 2025).

A common implication across these works is that steerability can be inference-time rather than weight-level. Parameter mixing, latent modulation, policy documents, activation vectors, and numerical cost knobs all serve as runtime control interfaces (Wang et al., 2024, Kobalczyk et al., 2024, Chakrabarti et al., 19 Dec 2025, Ghosh et al., 1 Jun 2025, Berant et al., 3 Dec 2025). This suggests that “policy” in steerable policies need not mean only an RL policy; it may denote any conditional decision rule whose behavior is explicitly reparameterizable after training.

5. Vision-language-action policies and embodied steerability

Robotic VLAs have made the distinction between goal conditioning and execution steering particularly explicit. "FineVLA" defines steerability as “the ability to execute the same high-level goal in different ways according to user-specified execution constraints,” including active arm, target object, initial and final configuration, contact and approach, trajectory and orientation, object interaction, failure and recovery, and body motion (Hu et al., 26 May 2026). The paper constructs FineVLA-Data from 972,247 trajectories across 10 datasets, compresses them to 47,159 representative trajectories, and annotates 220,606 steps with fine-grained descriptions. In policy training, it varies the ratio of fine-grained to raw goal-level instructions and finds a consistent inverted-U trend peaking at FG:Raw $$z \sim p_0(z)=\mathcal N(0,I), \qquad a = D_\theta(s,z),$$9 to $\pi_\phi^Z(z\mid s)=\mathcal N(\mu_\phi(s),\Sigma)$0. The best mixed setting reaches $\pi_\phi^Z(z\mid s)=\mathcal N(\mu_\phi(s),\Sigma)$1 in RoboTwin simulation and $\pi_\phi^Z(z\mid s)=\mathcal N(\mu_\phi(s),\Sigma)$2 in real-world dual-arm manipulation, compared with $\pi_\phi^Z(z\mid s)=\mathcal N(\mu_\phi(s),\Sigma)$3 for Raw-only; factor-specific gains are largest on pose $\pi_\phi^Z(z\mid s)=\mathcal N(\mu_\phi(s),\Sigma)$4, color $\pi_\phi^Z(z\mid s)=\mathcal N(\mu_\phi(s),\Sigma)$5, and approach direction $\pi_\phi^Z(z\mid s)=\mathcal N(\mu_\phi(s),\Sigma)$6 (Hu et al., 26 May 2026). FineVLA therefore treats steerability as instruction-conditioned control over execution details, not merely outcome success.

"Steerable Vision-Language-Action Policies for Embodied Reasoning and Hierarchical Control" expands the control interface further by training a low-level VLA on a spectrum of command abstractions: task-level descriptions, subtasks, atomic motions, gripper traces, grounded points, and hybrid commands, all encoded as text (Chen et al., 13 Feb 2026). This allows both a learned high-level embodied reasoner and an off-the-shelf VLM to choose the abstraction level most suitable for the current situation. With a learned high-level reasoner, the aggregate success rate reaches $\pi_\phi^Z(z\mid s)=\mathcal N(\mu_\phi(s),\Sigma)$7, compared with $\pi_\phi^Z(z\mid s)=\mathcal N(\mu_\phi(s),\Sigma)$8 for standard OpenVLA and $\pi_\phi^Z(z\mid s)=\mathcal N(\mu_\phi(s),\Sigma)$9 for ECoT (Chen et al., 13 Feb 2026). In an in-context hierarchical setting with an API VLM, the full method achieves $z=\mu_\phi(s)$0 average task progression versus $z=\mu_\phi(s)$1 for a SayCan-like subtask-only baseline (Chen et al., 13 Feb 2026). The paper’s central claim is that richer low-level steerability unlocks pretrained VLM reasoning that would otherwise be bottlenecked by a task-language-only interface.

ReSteer contributes a complementary diagnostic view. It defines steerability for a language-conditioned multitask robot policy in terms of whether a policy can switch from task $z=\mu_\phi(s)$2 to task $z=\mu_\phi(s)$3 at an intermediate state and still succeed. The paper formalizes the Steerability Coverage Ratio,

$z=\mu_\phi(s)$4

and shows that strong multitask performance does not imply strong task steerability (Chen et al., 18 Mar 2026). It then proposes a conditional mutual information proxy $z=\mu_\phi(s)$5, a steerability estimator for identifying low-steerability states, a steerable data generator, and a self-refinement pipeline. On LIBERO, ReSteer improves steerability by $z=\mu_\phi(s)$6 over 18k rollouts; in real-world experiments it raises steering success from about $z=\mu_\phi(s)$7 for a finetuned baseline to about $z=\mu_\phi(s)$8, a $z=\mu_\phi(s)$9 improvement (Chen et al., 18 Mar 2026). Together with FineVLA, this establishes that steerability in embodied systems is a measurable property of the closed-loop policy, not an automatic by-product of multitask pretraining.

6. Evaluation criteria, limitations, and open directions

The literature evaluates steerability with markedly different observables. In stochastic control and motion planning, the key metrics are terminal covariance, chance-constraint satisfaction, optimality gaps, and finite-time completeness or resolution-optimality (Kumagai et al., 29 Jan 2026, Bakolas, 2020, Fu et al., 2021, Fu et al., 2021). In robotic RL and manipulation, steerability is assessed via return, variance across seeds, path-tracking fidelity, factor-specific instruction adherence, task-switch success, and real-robot execution stability (Wang et al., 15 Jun 2026, Hu et al., 26 May 2026, Chen et al., 18 Mar 2026). In language systems, evaluation uses policy-conditioned F1, cost-penalized reward, precision-recall under alternative policies, and utility metrics such as helpfulness and coherence under activation steering (Chakrabarti et al., 19 Dec 2025, Berant et al., 3 Dec 2025, Ghosh et al., 1 Jun 2025). This diversity suggests that steerability is not a single metric but a property of controllable responsiveness under a specified steering interface.

Several limitations recur. Latent steering is limited by the support of the underlying generator or policy family; SteerGenPO explicitly notes that latent steering cannot create behaviors outside the action manifold discovered in Stage I (Wang et al., 15 Jun 2026). Covariance steering remains non-convex under chance constraints, and square-root formulations provide local rather than global guarantees in that regime (Kumagai et al., 29 Jan 2026). FineVLA and ReSteer both report persistent compositional generalization gaps, especially for unseen combinations such as arm-target bindings or arbitrary mid-execution task switches (Hu et al., 26 May 2026, Chen et al., 18 Mar 2026). SafeSteer reports strong empirical gains but does not provide systematic adversarial robustness analysis (Ghosh et al., 1 Jun 2025). CoPE is English-only and limited to semantic harm domains in its current evaluation (Chakrabarti et al., 19 Dec 2025). Steerable needle planning and related geometric methods assume static environments, kinematic models, and sufficiently accurate state estimation (Fu et al., 2021, Fu et al., 2021).

A plausible synthesis is that the field is converging on three design principles. First, steerability benefits from an explicit low-dimensional interface: latent vectors, reward weights, covariance targets, command abstractions, or policy text. Second, successful steering typically requires geometry in the control space: safety-preserving pullbacks, ordered latent manifolds, or task-conditioned action distributions aligned with value or semantics (Li et al., 29 Jan 2026, Wu et al., 20 May 2026, Wang et al., 2024). Third, steerability must be trained or engineered directly rather than assumed to emerge from scale alone; this is shown by contradictory-example training in content moderation, mixed fine-grained supervision in VLA learning, and data-generation or self-refinement pipelines for task-switchable robot control (Chakrabarti et al., 19 Dec 2025, Hu et al., 26 May 2026, Chen et al., 18 Mar 2026).

Steerable policies therefore designate not one method but a family of architectures and control formulations in which behavior remains adjustable after the base policy has been learned. The steering variable may act in latent space, task space, probability space, language space, or safety-constrained geometric space, but the objective is consistent: expose a structured interface through which behavior can be redirected predictably, efficiently, and, in some settings, with formal guarantees (Wang et al., 15 Jun 2026, Kumagai et al., 29 Jan 2026, Chen et al., 13 Feb 2026, Wu et al., 20 May 2026).