---
title: 'VelFlow: Embodiment-Agnostic Navigation Policy'
url: https://www.emergentmind.com/topics/velflow
type: topic
---

# VelFlow: Embodiment-Agnostic Navigation Policy

Searching arXiv for the specified paper to ground the article in the source.
VelFlow is the conditional normalizing-flow policy introduced in CE-Nav as the embodiment-agnostic General Expert for cross-embodiment local navigation. It models the full conditional distribution over body-velocity commands $a_t=(v_x,v_y,v_{\mathrm{yaw}})$ given the current observation $o_t$ and goal $g_t$, rather than predicting a single deterministic action. Within the CE-Nav two-stage pipeline, VelFlow is trained offline by imitation learning on planner-generated data and then frozen; a lightweight Dynamics-Aware Refiner is subsequently trained by online reinforcement learning to adapt to a target robot’s specific dynamics and controller imperfections. The design is intended to decouple universal geometric reasoning from embodiment-specific dynamic adaptation, while directly addressing the “disastrous averaging” failure mode of single-mode regression policies [2509.23203].

## 1. Position within CE-Nav

VelFlow appears in CE-Nav as the first-stage policy in a two-stage IL-then-RL framework. The first stage trains a general expert offline using imitation learning; the second stage freezes that expert and uses it as a guiding prior for a refiner trained online with reinforcement learning. In this decomposition, VelFlow is responsible for learning the distribution of kinematically-sound actions from a large-scale dataset generated by a classical planner, whereas the refiner compensates for robot-specific dynamics and controller imperfections [2509.23203].

This separation is central to the role of VelFlow. The offline expert is explicitly embodiment-agnostic: the training data are generated without any real-robot data and without embodiment-specific simulation. A plausible implication is that VelFlow is not intended to encode morphology-specific actuation idiosyncrasies; instead, it captures geometric navigation structure that remains useful across quadrupeds, bipeds, and quadrotors. The CE-Nav results further suggest that VelFlow functions primarily as a prior over feasible local actions, not as a complete end-to-end controller.

## 2. Probabilistic formulation

At each time step $t$, VelFlow models the conditional action density
$$
p(a_t \mid o_t, g_t),
$$
where $a_t=(v_x,v_y,v_{\mathrm{yaw}})$ is the body-velocity command and $o_t$ consists of LiDAR scan and proprioception. The model introduces an invertible function $f(\cdot; o_t,g_t)$ that maps action $a_t$ to a latent variable $z$ drawn from a simple base density $p_z(z)$, specified as a standard Gaussian. By the change-of-variables theorem,
$$
p(a_t\mid o_t,g_t)=p_z\bigl(z=f(a_t; o_t,g_t)\bigr)\cdot
\left|\det \frac{\partial f(a_t; o_t,g_t)}{\partial a_t}\right|^{-1}.
$$

This formulation gives VelFlow exact likelihood evaluation under an invertible transformation. The paper’s conclusion attributes three properties to this conditional-flow design: exact log-likelihood training, single-pass multi-modal sampling, and tractable gradients [2509.23203]. In context, these properties distinguish VelFlow from deterministic regression guides, which collapse multiple valid decisions into an average, and from alternatives such as diffusion policies, which the paper reports as slower at inference.

## 3. Architecture and conditioning mechanism

VelFlow is built as a sequence of 12 affine coupling blocks in the style of Real-NVP, with intermediate 512-dimensional hidden nets. Between every few coupling blocks, the architecture inserts an invertible linear “1×1” rotation and an elementwise ActNorm layer for stability. The state-conditioning pathway uses a state encoder that processes the 360° LiDAR scan, represented as 144 rays through a three-layer CNN, concatenated with the 7D robot state—goal direction, velocities, and distance—through a 2-layer MLP to produce a 256-dimensional embedding $s_t$ [2509.23203].

For the explicit coupling-layer form, let $D=3$ for $(v_x,v_y,v_{\mathrm{yaw}})$ and choose $d=\lfloor D/2\rfloor=1$. Writing
$$
a_t=[a_1; a_{2:3}],
$$
each coupling layer computes
$$
y_1=a_1,
$$
$$
y_{2:3}=a_{2:3}\cdot \exp\bigl(s(a_1,s_t)\bigr)+t(a_1,s_t).
$$
Here $s(\cdot)$ and $t(\cdot)$ are small MLPs with 512 hidden units, each conditioned on the 256-D state embedding. The full transformation $f(\cdot)$ is the composition of 12 such blocks plus linear and ActNorm layers.

The conditioning mechanism is therefore localized at each coupling layer: the scale and translate subnetworks take as input the “frozen” half of the split action and the state embedding $s_t$. This suggests that VelFlow is organized to preserve invertibility in the action channel while allowing rich conditioning on perception and goal state.

## 4. Offline data generation and maximum-likelihood training

VelFlow is trained entirely from expert data generated by a classical planner. The data-generation procedure uses the Dynamic Window Approach in a purely 2D geometric simulator with a circular agent and generic dynamic bounds. In each procedurally generated obstacle environment, the dataset includes all candidate velocity commands whose DWA objective score satisfies
$$
\mathrm{score} \ge (1-\delta)\cdot \mathrm{Score}_{\max}, \qquad \delta=0.1.
$$
This collection rule is intended to retain multiple high-scoring actions, such as left-versus-right decisions around an obstacle. After discarding failed trajectories, the final dataset comprises 10 million successful $(o_t,a_t)$ pairs [2509.23203].

Training uses maximum likelihood through the expected negative log-likelihood over the expert dataset:
$$
\mathcal{L}_{\mathrm{NLL}}
=
-
\mathbb{E}_{(o_t,a_t)\sim \mathcal{D}_{\mathrm{expert}}}
\bigl[\log p(a_t\mid o_t,g_t)\bigr].
$$

No additional data augmentation is applied, because the DWA-generated set already contains multi-modal examples. ActNorm layers and weight-tying in the flows serve as implicit regularizers. In methodological terms, the training setup is notable for what it excludes: no real-robot expert data and no embodiment-specific simulation are used in constructing the general expert. A plausible implication is that the learned prior is constrained to the geometric and kinematic structure recoverable from the planner’s action distribution.

## 5. Multi-modality and the “disastrous averaging” problem

A central motivation for VelFlow is the claim that deterministic regression fails when multiple actions are equally valid. The paper characterizes this failure as “disastrous averaging”: if “turn-left” and “turn-right” are both acceptable, a single-mode regression network tends to learn their average, which can correspond to driving straight into an obstacle [2509.23203].

Because VelFlow models the full conditional density $p(a_t\mid o_t,g_t)$, it can generate diverse action proposals by sampling $z\sim \mathcal{N}(0,I)$ and applying the inverse transform
$$
a=f^{-1}(z; o_t,g_t).
$$
The resulting samples are described as naturally diverse and clusterable. In the paper’s Fig. 6, 100 rollouts approaching a T-junction split into two clusters: the expert’s samples cluster around left and right turns, and the refiner’s final commands preserve this bimodality while adapting to dynamic constraints.

The reported quantitative comparison reinforces this interpretation. Under 300 obstacles, the MLP-regression guide yields $\mathrm{SR}\approx 0.46$, whereas the VelFlow guide yields $\mathrm{SR}\approx 0.84$. The same ablation sequence reports that replacing VelFlow with a diffusion policy improves over regression but still trails VelFlow and is 8× slower at inference. Taken together, these results position VelFlow as the mechanism by which CE-Nav represents multi-modal local decisions without collapsing them into a single averaged command.

## 6. Empirical behavior, transfer, and limitations

The paper distinguishes sharply between VelFlow as an offline expert and CE-Nav as the complete two-stage system. VelFlow alone, used as the General Expert without the reinforcement-learning refiner, fails to track and reaches $\mathrm{SR}\approx 0.01$ at high density due to covariate shift. By contrast, pure RL without guidance requires approximately 52 h of training versus 6 h with VelFlow guidance. This is an important correction to a possible misconception: VelFlow is not presented as sufficient by itself for robust deployment under target-robot dynamics; its role is to provide the guiding prior that makes rapid adaptation feasible [2509.23203].

After a 6 h RL refiner stage, CE-Nav achieves cross-embodiment transfer results reported as follows.

| Setting | Reported result |
|---|---|
| Three quadrupeds | $mSR\sim 0.85$–$0.86$ |
| Biped | $0.75$ |
| Quadrotor | $0.80$ |

On Go2, the comparison to baselines is reported more explicitly.

| Method | mSR | mSPL |
|---|---:|---:|
| CE-Nav | 0.8575 | 0.82 |
| DWA | 0.64 | 0.60 |
| NavRL | 0.69 | 0.65 |
| BC | 0.03 | 0.025 |
| DP | 0.07 | 0.064 |

The ablation results also indicate that static high weighting $\lambda=0.5$ under-explores, whereas an annealed curriculum yields the best SR+SPL balance. This suggests that the usefulness of VelFlow as a prior depends not only on its density model but also on how strongly the second-stage learner is constrained by that prior during adaptation.

In aggregate, the empirical picture is internally consistent. VelFlow captures the planner-induced distribution of kinematically-sound actions; by itself it remains vulnerable to covariate shift and tracking mismatch, but when frozen and paired with a lightweight, curriculum-guided RL refiner it supports rapid adaptation across substantially different robot morphologies. The paper therefore presents VelFlow less as a standalone navigation policy than as a robust, embodiment-agnostic probabilistic prior over local velocity commands.

Source: https://www.emergentmind.com/topics/velflow