PF2MP: Potential Field-Guided Flow Policy

• PF2MP is a multimodal flow-matching policy that integrates a learned flow field with an auxiliary guidance signal derived from scalar potentials.
• It leverages streamable action-space substrates like SFP and explicit mechanisms—such as critic gradients and force-informed cues—to bias trajectories.
• This approach balances multimodal generation with targeted guidance, offering practical benefits for safe trajectory planning and contact-rich manipulation.

Searching arXiv for papers relevant to PF2MP and closely related flow-matching policy methods. Potential Field-Guided Flow Matching Policy (PF2MP) denotes an interpretive class of methods that combine a flow-matching policy with an additional guidance signal that shapes the generated vector field in action space or trajectory space. The exact term is not standardized in the cited literature, but several papers provide the constituent ingredients: streamable action-space flow policies, trajectory-level motion fields, critic- or objective-gradient guidance, and force- or compliance-informed directional shaping. Across these works, the common pattern is a learned flow prior that generates multimodal actions or paths, together with an auxiliary field that biases generation toward desirable regions such as demonstrations, higher-value actions, task objectives, or contact-maintaining motions (Jiang et al., 28 May 2025).

1. Conceptual definition and scope

PF2MP is best understood as an overview of several neighboring ideas rather than as a single canonical algorithm. The strongest recurring structure is a learned flow field that maps a simple source distribution to an action or trajectory distribution, plus an explicit or implicit guidance mechanism that alters local motion geometry during training, inference, or both. In this sense, the “flow matching policy” part refers to a policy parameterized by a velocity field or transport map, while the “potential field-guided” part refers to an added directional bias derived from a scalar objective, cost, critic, force cue, or task-specific field (Nguyen et al., 8 Mar 2025).

A central distinction in this literature is between methods that guide the flow directly and methods that only use a scalar field indirectly. Some approaches introduce an explicit guidance gradient into the generative dynamics, such as objective-guided posterior correction in motion planning or critic-gradient steering in reinforcement learning. Others instead use scalar fields to reweight examples, regularize a student actor, or shape a rollout controller after the flow policy has already produced a trajectory. This suggests that PF2MP is not a single mechanism but a family of architectures spanning direct vector-field modification, loss shaping, reranking, and execution-time compliance shaping (Shi et al., 21 Jun 2026).

The term “potential field” is also broader here than in classical artificial potential field robotics. In the cited work, the scalar guidance source may be a demonstration-tracking quadratic term, a learned action-value function, a task objective over trajectories, an FM-derived state-action energy surrogate, or a force-informed virtual target. A plausible implication is that PF2MP should be defined functionally—by how it shapes the learned flow—rather than by requiring a specific conservative field construction (Li et al., 3 Oct 2025).

2. Flow-matching policy substrate

The most direct action-space substrate for PF2MP is the Streaming Flow Policy (SFP), which collapses the latent generative trajectory and the executed action trajectory into a single neural ODE in action space (Jiang et al., 28 May 2025). In SFP, history-conditioned imitation learning is posed with observation history $h=\{o_i\}_{i=1}^K$, future action trajectory $\xi:[0,1]\to A$, data distribution $p_D(h,\xi)$, and learned velocity field

$$v_\theta(a,t\mid h), \qquad \frac{da}{dt}=v_\theta(a,t\mid h).$$

Instead of sampling an entire action sequence from pure Gaussian noise in $A^T$, SFP initializes from a narrow Gaussian around the previous action $a_{\mathrm{prev}}$: $$a(t)=a_0+\int_0^t v_\theta(a(s),s\mid h)\,ds,\qquad a_0\sim\mathcal N(a_{\mathrm{prev}},\sigma_0^2).$$ This makes the evolving state $a(t)$ itself executable, so flow time coincides with action-execution time (Jiang et al., 28 May 2025).

Its conditional flow construction is especially important for PF2MP: $$v_\xi(a,t)=\dot{\xi}(t)-k(a-\xi(t)),\qquad p_\xi^0(a)=\mathcal N(a\mid \xi(0),\sigma_0^2).$$ The first term follows the demonstration tangent, while the second is a stabilizing attraction term toward the demonstration at the same time index. The induced conditional marginal is

$$p_\xi(a\mid t)=\mathcal N(a\mid \xi(t),\sigma_0^2 e^{-2kt}),$$

so deviations contract exponentially. The paper does not phrase this in potential-field language, but it is mathematically natural to interpret $\xi:[0,1]\to A$0 as the negative gradient of a quadratic local potential centered on the demonstration. This is the clearest demonstration-centered prototype for PF2MP-style local guidance (Jiang et al., 28 May 2025).

The learning objective is conditional flow matching: $\xi:[0,1]\to A$1 The corresponding optimal field matches per-timestep marginal action distributions rather than the full joint trajectory distribution. That distinction is central: it preserves multimodality at each time slice but can compose segments from different demonstrations. This suggests that PF2MP built on such a substrate will inherit both the compositionality and the marginal-only guarantee (Jiang et al., 28 May 2025).

3. Guidance mechanisms as potential-like fields

Several recent methods instantiate guidance more explicitly by adding objective or value gradients to a learned flow prior. FlowMP is the clearest planning-side example. It learns a continuous motion field $\xi:[0,1]\to A$2 over trajectories, including position-, velocity-, and acceleration-related components, and introduces objective-guided posterior correction at inference (Nguyen et al., 8 Mar 2025). Its posterior is written as

$\xi:[0,1]\to A$3

with

$\xi:[0,1]\to A$4

leading to MAP guidance

$\xi:[0,1]\to A$5

The guidance gradient is

$\xi:[0,1]\to A$6

and the posterior flow is described as

$\xi:[0,1]\to A$7

Although not stated as a potential field, this is structurally equivalent to adding the gradient of a trajectory-level energy to a learned flow prior (Nguyen et al., 8 Mar 2025).

On the policy-improvement side, FlowDPG provides a local action-space guidance construction for flow-matching policies. Starting from an intermediate noisy point $\xi:[0,1]\to A$8, it defines a projected clean action

$\xi:[0,1]\to A$9

computes a critic gradient

$p_D(h,\xi)$0

normalizes it relative to the demonstration transport magnitude,

$p_D(h,\xi)$1

and forms a critic-improved target

$p_D(h,\xi)$2

The actor then regresses toward a mixture of the demonstration velocity $p_D(h,\xi)$3 and the critic-guided velocity $p_D(h,\xi)$4. This gives a concrete recipe for a PF2MP-style policy: preserve a feasible demonstration-driven transport and add a bounded local correction from a scalar field gradient (Shi et al., 21 Jun 2026).

Q-VGM extends the same idea to flow-matching vision-language-action policies by explicitly turning a learned scalar value function into a denoising-time residual velocity field. Its central mapping is: from denoising state $p_D(h,\xi)$5, form a base clean-action estimate

$p_D(h,\xi)$6

optimize that clean action by projected ascent on $p_D(h,\xi)$7,

$p_D(h,\xi)$8

then transport the endpoint displacement back into a denoising-time vector field: $p_D(h,\xi)$9 The policy is updated by matching the residual flow velocity to this target. This is one of the most direct algorithmic realizations of a potential-guided flow policy in the cited literature, with the learned $$v_\theta(a,t\mid h), \qquad \frac{da}{dt}=v_\theta(a,t\mid h).$$0 functioning as the scalar potential (Wang et al., 6 Jun 2026).

4. Training-time shaping, regularization, and teacher guidance

Not all PF2MP-relevant methods guide the flow by injecting gradients into the dynamics. Several instead shape the learned distribution through weighting, regularization, or a teacher field. Guided Flow Policy (GFP) is a representative case in offline RL. It combines a multi-step flow-matching policy, denoted Value-aware Behavior Cloning (VaBC), with a distilled one-step actor. The flow regularizer is trained by weighted flow matching: $$v_\theta(a,t\mid h), \qquad \frac{da}{dt}=v_\theta(a,t\mid h).$$1 where

$$v_\theta(a,t\mid h), \qquad \frac{da}{dt}=v_\theta(a,t\mid h).$$2

This does not directly modify the flow ODE with a gradient field, but it does reshape the action distribution toward higher-value dataset actions. A plausible implication is that PF2MP can be realized as value-weighted density shaping even when no explicit force-like term is added to the generative dynamics (Tiofack et al., 3 Dec 2025).

FM-IRL uses a different decomposition: a teacher FM model over joint state-action space induces a scalar compatibility landscape

$$v_\theta(a,t\mid h), \qquad \frac{da}{dt}=v_\theta(a,t\mid h).$$3

which is converted into an FM-enhanced discriminator

$$v_\theta(a,t\mid h), \qquad \frac{da}{dt}=v_\theta(a,t\mid h).$$4

and reward

$$v_\theta(a,t\mid h), \qquad \frac{da}{dt}=v_\theta(a,t\mid h).$$5

The student policy is regularized toward teacher-generated expert-like actions via

$$v_\theta(a,t\mid h), \qquad \frac{da}{dt}=v_\theta(a,t\mid h).$$6

The paper does not frame this as a potential field, but its $$v_\theta(a,t\mid h), \qquad \frac{da}{dt}=v_\theta(a,t\mid h).$$7 acts as a scalar energy surrogate over state-action space, and the student regularizer acts as an attraction toward an expert-supported manifold (Wan et al., 10 Oct 2025).

FlowMPC illustrates another training-versus-inference separation. It leaves the imitation-trained flow policy unchanged and instead uses a learned world model with MPPI at test time to score candidate action trajectories. The central planning score is

$$v_\theta(a,t\mid h), \qquad \frac{da}{dt}=v_\theta(a,t\mid h).$$8

The paper is explicit that this is external, plug-and-play guidance rather than retraining the flow objective. For PF2MP, this suggests a modular design in which a fixed flow policy provides multimodal proposals and a potential-derived score reranks or refines them at inference (Hamel, 15 Jun 2026).

5. Execution architectures: streaming, one-step maps, and best-of-$$v_\theta(a,t\mid h), \qquad \frac{da}{dt}=v_\theta(a,t\mid h).$$9

PF2MP is also shaped by how the learned flow is executed. One major axis is whether the policy is streamed, one-step distilled, or used as a multi-step stochastic planner. SFP is explicitly streamable: during receding-horizon execution it integrates only over the current chunk and sends each intermediate action to the robot before the chunk is complete. Its Euler update is

$A^T$0

Because intermediate states lie in action space rather than noisy trajectory space, they are directly executable. This streaming formulation is especially compatible with online local guidance fields, since vector evaluations and control execution overlap naturally (Jiang et al., 28 May 2025).

Flow map policies go further by learning arbitrary-size jumps across generative time rather than only instantaneous velocities. A flow map $A^T$1 satisfies

$A^T$2

with

$A^T$3

The special case $A^T$4 yields one-step generation from noise to action. The paper then derives a trust-region-optimal closed-form guidance target

$A^T$5

which is regressed by the online actor. This is directly relevant to PF2MP because it shows that guidance need not be applied by expensive iterative flow integration; it can instead be distilled into an arbitrary-step transport map (Ziakas et al., 12 May 2026).

At the planner level, Flow Motion Policy (FMP) uses flow matching to generate multimodal joint-increment path chunks and exploit best-of-$A^T$6 sampling. The learned conditional field is trained on

$A^T$7

with inference

$A^T$8

Best-of-$A^T$9 sampling evaluates candidate paths by

$a_{\mathrm{prev}}$0

This is not potential guidance in the strict sense, but it is a practical substrate for PF2MP: the multimodal flow prior is already present, and a potential-based reranker or drift term could replace or augment the post hoc collision count (Soleymanzadeh et al., 8 Apr 2026).

6. Contact-rich and physically grounded variants

A particularly important branch of PF2MP-like reasoning appears in contact-rich manipulation, where guidance signals are force- and compliance-related rather than purely geometric. “Flow with the Force Field” provides the strongest example. It learns a conditional flow-matching policy over outputs

$a_{\mathrm{prev}}$1

conditioned on

$a_{\mathrm{prev}}$2

where $a_{\mathrm{prev}}$3 is a point cloud, $a_{\mathrm{prev}}$4 the end-effector pose, and $a_{\mathrm{prev}}$5 the force input. The flow ODE is

$a_{\mathrm{prev}}$6

with standard flow-matching regression toward $a_{\mathrm{prev}}$7 (Li et al., 3 Oct 2025).

Its guidance mechanism does not take the form of an explicit scalar potential. Instead, it uses a force-informed virtual target during synthetic data generation: $a_{\mathrm{prev}}$8 The authors describe this as analogous to “a target pose inside the object,” so that moving toward the target generates force on the object surface. At execution time, the predicted reference trajectory and virtual target are converted into a compliant direction field: $a_{\mathrm{prev}}$9

$$a(t)=a_0+\int_0^t v_\theta(a(s),s\mid h)\,ds,\qquad a_0\sim\mathcal N(a_{\mathrm{prev}},\sigma_0^2).$$0

This field is executed through a passive impedance controller. Although there is no explicit $$a(t)=a_0+\int_0^t v_\theta(a(s),s\mid h)\,ds,\qquad a_0\sim\mathcal N(a_{\mathrm{prev}},\sigma_0^2).$$1 with $$a(t)=a_0+\int_0^t v_\theta(a(s),s\mid h)\,ds,\qquad a_0\sim\mathcal N(a_{\mathrm{prev}},\sigma_0^2).$$2, the construction is unmistakably PF2MP-like: a learned flow policy produces trajectories and compliance parameters, while a local physically grounded field biases motion toward contact-maintaining directions (Li et al., 3 Oct 2025).

The empirical pattern in that work is also notable. For block flipping, the full point-cloud + force + compliance model substantially outperforms versions without force or without compliant output. A plausible implication is that PF2MP methods in contact-rich settings may benefit more from physically meaningful guidance fields than from purely geometric action-generation priors (Li et al., 3 Oct 2025).

7. Limitations, misconceptions, and open directions

A common misconception is that any flow-matching policy with a stabilizing term already constitutes a full potential-field-guided policy. The literature does not support that equivalence. SFP’s demonstration-attracting term is mathematically analogous to a local quadratic attractor, but the paper does not introduce obstacle potentials, barrier functions, Lyapunov guarantees, or safety certificates. It guarantees only per-timestep marginal matching, not full joint trajectory fidelity, and its current execution freezes observations within a chunk (Jiang et al., 28 May 2025).

Another misconception is that value guidance and potential-field guidance are interchangeable without qualification. Critic-guided methods such as FlowDPG, Q-VGM, and FMQ use gradients of learned scalar functions over actions, but these functions are task-value surrogates rather than explicit geometric or safety potentials. Their gradients can be effective steering signals, yet their reliability depends on critic quality, projection consistency, and support of the training distribution. Several papers emphasize stabilizers such as consistency regularization, bounded guidance magnitude, critic ensembling, clipped objectives, and conservative targets precisely because unguided scalar gradients can be unreliable off-distribution (Wang et al., 6 Jun 2026).

The execution-time trade-off between multimodality and explicit guidance also remains unresolved. Flow priors naturally support multiple modes, while classical potential fields often collapse toward a single attractor or suffer from local minima. Methods such as FlowMPC, FMP best-of-$$a(t)=a_0+\int_0^t v_\theta(a(s),s\mid h)\,ds,\qquad a_0\sim\mathcal N(a_{\mathrm{prev}},\sigma_0^2).$$3, and Q-guided beam search suggest one path forward: keep the multimodal generative policy intact and apply external guidance through reranking, beam selection, or bounded local refinement rather than replacing the flow with a purely deterministic field (Ziakas et al., 12 May 2026).

Across the cited work, several open directions emerge. One is to define scalar potentials directly over denoising or flow-map states rather than only over clean actions. Another is to combine explicit task or obstacle potentials with second-order or acceleration-aware flow models, as suggested by FlowMP’s motion-field formulation. A third is to integrate explicit safety shaping into streaming or one-step flow-map execution so that low-latency control does not rely solely on data-induced regularities. These directions are strongly suggested by the current literature, but they remain extrapolations rather than established results (Nguyen et al., 8 Mar 2025).

In this sense, PF2MP presently names a coherent research direction rather than a single accepted algorithm: a multimodal flow-based policy whose local geometry is shaped by a guidance field, with implementations ranging from demonstration-tube attractors and critic-gradient steering to world-model reranking and force-informed compliant execution.