Skip to content

ElasticFlow: One-Step Physics-Consistent Policy with Elastic Time Horizons for Language-Guided Manipulation

Conference: ACL 2026
arXiv: 2605.08799
Code: To be confirmed
Area: Embodied AI / Diffusion Policy / Flow Matching / Robot Manipulation
Keywords: One-step diffusion, Mean velocity field, Elastic time, Flow matching, VLA

TL;DR

The authors propose ElasticFlow, which replaces instantaneous velocity field learning with a mean velocity field (MeanFlow) to learn language-conditioned robot actions. By explicitly encoding control granularity via an "Elastic Time Horizon \(\Delta t=t-r\)," it achieves 1-NFE single-step inference (~71Hz) and outperforms OpenVLA and \(\pi_0\) on long-horizon tasks such as LIBERO-Long and CALVIN ABC-D.

Background & Motivation

Background: In Embodied AI, generalist policies that map visual observations and language instructions to continuous actions primarily follow two paths: Diffusion Policies (e.g., Diffusion Policy, \(\pi_0\)) have become mainstream due to their strong multimodal modeling capabilities; autoregressive VLAs (e.g., OpenVLA, RT-2) rely on token discretization for language-action alignment.

Limitations of Prior Work: Iterative denoising in diffusion policies requires dozens of NFEs (Network Function Evaluations), leading to latency > 100ms and control frequencies of only 8–12Hz, which cannot respond to rapidly changing physical environments (e.g., intercepting rolling objects). Existing acceleration schemes (Consistency Model, Progressive Distillation) require complex teacher-student pipelines and often sacrifice "physics consistency," leading to high-frequency jitter (jerk) or non-smooth paths. Autoregressive VLAs are even slower (~5Hz) due to token-by-token generation, and discretization introduces quantization errors.

Key Challenge: (1) There is a trade-off between inference speed and physics consistency; simply reducing steps causes trajectories to lose geometric smoothness. (2) The "temporal heterogeneity" of robot tasks is often ignored—short-range reactive control requires millisecond jitter suppression, while long-range tasks require second-level trajectory planning. Traditional fixed-horizon networks fall into Spectral Bias, failing to model high-frequency and low-frequency signals simultaneously.

Goal: (1) Achieve 1-NFE inference without distillation. (2) Ensure single-step predictions remain geometrically smooth and physically consistent. (3) Enable a single set of weights to perform both millisecond-level reactive control and long-range multi-stage planning.

Key Insight: MeanFlow, proposed by Geng et al. (2025) in generative modeling, directly learns the "mean velocity" over a time interval \([r,t]\) rather than the instantaneous velocity. A single forward pass yields the overall displacement from noise to data. The authors adapt this to robot action flows and expose \(\Delta t=t-r\) to the network as a "control granularity knob."

Core Idea: The method utilizes a mean velocity field \(u(z_t,r,t)\) instead of instantaneous velocity \(v(z_t,t)\) for action generation, using \(\Delta t\) as a "Spectral Zoom Lens" to unify short-range reaction and long-range planning.

Method

Overall Architecture

Observations \(o\) are encoded via SigLIP, and language instructions \(\ell\) are encoded via T5. These are injected into a DiT backbone (150M parameters) via cross-attention. The Elastic Time Horizon module encodes \((r,t,\Delta t)\) using Fourier features, which are injected via AdaLN modulation. The output is the mean velocity field prediction \(u_\theta(z_t,r,t,o,\ell)\). Training is supervised using the MeanFlow Identity Loss with forward-mode AD. During inference, given \(z_1\sim\mathcal{N}(0,I)\), a single forward pass calculates the action chunk \(\hat{x}=z_1-u_\theta(z_1,0,1,o,\ell)\) without any iteration or distillation.

Key Designs

  1. Mean Velocity Field Modeling (MeanFlow Identity):

    • Function: Transforms action generation from "solving multi-step ODE integration" into "learning a one-step mapping," while implicitly incorporating trajectory curvature correction to ensure physical smoothness.
    • Mechanism: Define \(u(z_t,r,t)\triangleq\frac{1}{t-r}\int_{r}^{t}v(z_\tau,\tau)d\tau\). From the fundamental theorem of calculus, the identity \(u(z_t,r,t)=v(z_t,t)-(t-r)\frac{d}{dt}u(z_t,r,t)\) is derived, where \(\frac{d}{dt}\) is the total derivative (including \(v\cdot\nabla_z u\) and \(\partial_t u\)). The training objective aligns the network prediction with the right side of this identity. The instantaneous velocity ground truth is constructed using optimal transport paths \(v(z_t,t)=x_{\text{target}}-x_{\text{noise}}\).
    • Design Motivation: Instantaneous velocity only describes local tangents, requiring ODE integration to recover global displacement, which makes multi-step inference a "necessary cost." By using mean velocity, a single step captures geometric information across the entire time interval, and the \((t-r)\frac{d}{dt}u\) term naturally acts as a "manifold curvature correction," suppressing high-frequency jitter at the formula level.

  2. Elastic Time Horizon:

    • Function: Uses a continuous parameter \(\Delta t=t-r\) to allow the same network to switch seamlessly between "short-range high-frequency reaction" and "long-range low-frequency planning."
    • Mechanism: In addition to absolute time \(t\), \(\Delta t\) is fed into the network. Both are encoded via Gaussian Fourier features \(\text{Emb}(r,t)=\text{MLP}([\text{FF}(t),\text{FF}(t-r)])\) and modulate the DiT via AdaLN. During inference, \(\Delta t\) is selected based on task granularity: small \(\Delta t\) focuses on local pose adjustment, while large \(\Delta t\) performs long-range trajectory planning, switching dynamically within a single weight space. During execution, the continuous flow is discretized into \(N\) steps based on the target control frequency, with a physical step size \(\delta t=T/N\).
    • Design Motivation: Neural networks exhibit Spectral Bias, making it difficult to fit high-frequency signals. Explicitly injecting \(\Delta t\) is equivalent to telling the network "which scale to observe," acting as a spectral zoom lens that covers both short-range transients and long-range structures under one set of weights. Mismatch tests (forcing an incorrect \(\Delta t\)) verify this physical significance.

  3. Forward-mode AD + Stop-gradient + CFG Training:

    • Function: Stably utilizes the MeanFlow Identity as a training objective while supporting language-conditioned Classifier-Free Guidance (CFG).
    • Mechanism: The loss is \(\mathcal{L}(\theta)=\mathbb{E}_{t,r,x_1,\epsilon,c}[\|u_\theta(z_t,r,t,o,c)-\text{sg}(\mathcal{T}_{\text{target}})\|_2^2]\), where \(\mathcal{T}_{\text{target}}=v(z_t,t)-(t-r)(v(z_t,t)\cdot\nabla_z u_\theta+\partial_t u_\theta)\). The Jacobian-vector product \(\nabla_z u_\theta\) is computed efficiently using forward-mode automatic differentiation (AD) to avoid Hessian overhead. Conditions \(c\in\{\ell,\emptyset\}\) are substituted with an empty set with probability \(p_{\text{drop}}\) for joint training. Inference uses \(\hat{x}=z_1-(u_\theta(\cdot,\emptyset)+w(u_\theta(\cdot,\ell)-u_\theta(\cdot,\emptyset)))\).
    • Design Motivation: Directly calculating gradients for bootstrapped targets can diverge; stop-gradient isolates self-referencing terms to stabilize optimization. Forward-mode AD addresses the computational bottleneck of second-order terms. CFG allows single-step inference to still adjust semantic alignment strength.

Loss & Training

The model minimizes the Mean Square Error (MSE) towards the target \(\mathcal{T}_{\text{target}}\) derived from the MeanFlow Identity. The stop-gradient \(\text{sg}(\cdot)\) stabilizes bootstrap optimization. Using a CFG drop probability \(p_{\text{drop}}\), the guidance scale \(w\) remains stable within \([1.5, 2.5]\) during inference. The 150M DiT backbone is lighter than the 300M UNet used in Diffusion Policy.

Key Experimental Results

Main Results

LIBERO Benchmarks + CALVIN ABC-D:

Benchmark / Metric ElasticFlow Prev. SOTA Key Baseline
LIBERO-Spatial (SR ↑) 98.4 98.8 (HiF-VLA) \(\pi_0\) 96.8
LIBERO-Object (SR ↑) 99.3 99.4 (HiF-VLA) \(\pi_0\) 98.8
LIBERO-Goal (SR ↑) 98.7 97.9 (OpenVLA-OFT) \(\pi_0\) 95.8
LIBERO-Long (SR ↑) 97.6 96.4 (HiF-VLA) \(\pi_0\) 85.2
LIBERO Average 98.5 98.0 (HiF-VLA) Octo 75.1
CALVIN ABC-D 3rd Person (Avg.Len. ↑) 4.15 4.08 (HiF-VLA) \(\pi_0\) 3.65
CALVIN ABC-D Multi-view (Avg.Len. ↑) 4.37 4.35 (HiF-VLA) \(\pi_0\) 3.92
RoboTwin Long & Extra Long (SR ↑) 71.1 69.0 (SimpleVLA-RL) \(\pi_0\) 43.3 / RDT 27.8

The 12 percentage point jump in LIBERO-Long (from 85.2 for \(\pi_0\) to 97.6) is the most significant result, as long-horizon tasks are where diffusion policies typically fail.

Inference Latency (RTX 4090, batch=1):

Method NFE Latency (ms) ↓ Frequency (Hz) ↑
OpenVLA (7B Transformer) Auto-reg. 200.0 ∼5
Diffusion Policy (300M UNet) 16 (DDIM) 120.0 ∼8
\(\pi_0\) (300M DiT) 10 (Euler) 85.0 ∼12
Consistency Policy 2 28.0 ∼35
ElasticFlow (150M DiT) 1 14.0 71

This represents a 5× speedup compared to Diffusion Policy and a 14× speedup compared to OpenVLA.

Ablation Study

Configuration Long Horizon SR Short Horizon SR Description
w/o Horizon Input 52.7% 61.5% Removing \(\Delta t\) loses 18.4 pts
Fixed \(\Delta t=10\) 58.2% 94.5% Strong short-range, weak long-range
Fixed \(\Delta t=50\) 62.1% 55.4% Slightly stronger long-range, short-range fails
Mismatch Force \(\Delta t=10\) on Long 45.3% — "Myopic": Only looks at immediate surroundings
Mismatch Force \(\Delta t=50\) on Short — 55.7% "Sluggish": Cannot react quickly
ElasticFlow (Dynamic \(\Delta t\)) 71.1% 98.2% Full model
Training Objective Step Success Rate Jerk ↓ Latency
Standard CFM (\(v_t\)) 1-NFE 12.4% \(8.5\times 10^{-2}\) 14ms
Standard CFM (\(v_t\)) 10-NFE 68.5% \(3.2\times 10^{-3}\) 140ms
ElasticFlow (\(u_t\)) 1-NFE 71.1% \(\mathbf{1.1\times 10^{-3}}\) 14ms

Key Findings

  • The \(\Delta t\) module is the primary contributor to ElasticFlow: removing it results in an 18.4 point drop. Mismatch tests (forcing the wrong \(\Delta t\)) caused success rates to fall to 45.3% for long-horizon and 55.7% for short-horizon tasks, disproving the feasibility of a "fixed horizon" approach.
  • The primary benefit of mean velocity field modeling is maintaining high physics consistency at 1-NFE: the Jerk of single-step ElasticFlow (\(1.1\times 10^{-3}\)) is lower than that of 10-step CFM (\(3.2\times 10^{-3}\)), proving that curvature correction terms suppress jitter at the mathematical level.
  • Standard Flow Matching fails in a single step (SR 12.4%), but ElasticFlow achieves 71.1%, indicating that speed is not simply a trade-off between model size and step count; the underlying modeling target must change.
  • ElasticFlow maintains 83.6% / 72.7% success rates on the 4th and 5th instructions in CALVIN, demonstrating superior stability in long instruction chains compared to baselines that decay after initial tasks.
  • The CFG guidance scale \(w\) is stable within \([1.5, 2.5]\) and insensitive to hyperparameters.

Highlights & Insights

  • "Learning average velocity instead of instantaneous velocity" represents a paradigm shift: the ODE integration bottleneck is moved from "runtime computation" to "training-time representation," making 1-NFE a mathematical equivalence rather than an engineering compromise.
  • The Elastic Time Horizon \(\Delta t\) is an extremely economical design: by adding just one dimensional parameter + Fourier encoding, a single network handles both high-frequency reaction and long-range planning, eliminating the need for cascaded networks of different horizons.
  • The "Spectral Zoom Lens" metaphor is more than just rhetorical—mismatch tests provide physical validation through counterfactual analysis, grounding the abstract concept of spectral bias in observable robot behavior.
  • The distillation-free nature provides significant engineering value: unlike Consistency Policy or Progressive Distillation which require a teacher-student pipeline, ElasticFlow achieves 1-NFE via one-stage training, reducing deployment costs.

Limitations & Future Work

  • The data scale is limited to millions of interactions and has not been verified against scaling laws on billion-level cross-embodiment data like Open X-Embodiment.
  • While forward-mode AD for Jacobian-vector products is cheaper than Hessian calculations, it still incurs significant overhead compared to standard training; a detailed wall-clock comparison was not provided.
  • The selection of \(\Delta t\) currently relies on task priors; ideally, \(\Delta t\) should be adaptively selected by the policy based on context.
  • Language conditions are only injected via cross-attention; semantic parsing of complex multi-stage instructions is still limited by the T5 encoder.
  • High-frequency (71Hz) single-step inference leaves very little time for error recovery; integration with MPC or online RL for closed-loop correction is a natural progression.
  • Sim-to-Real evaluation was limited to qualitative and partial quantitative tests on xArm6; reliability for cross-platform real-world deployment requires further validation.
  • vs. Diffusion Policy (Chi 2023) / \(\pi_0\): Both are flow/diffusion-based, but DP/\(\pi_0\) learn instantaneous velocity requiring multi-step integration; ElasticFlow learns the mean velocity field for 1-step, smoother physics.
  • vs. Consistency Policy (Prasad 2024) / Progressive Distillation: Those require teacher-student pipelines that are complex and often sacrifice diversity; ElasticFlow offers distillation-free one-stage training.
  • vs. OpenVLA / CogACT / OpenVLA-OFT: Autoregressive VLAs use token discretization, resulting in low frequency and quantization errors; ElasticFlow generates in continuous action space with an order of magnitude higher frequency.
  • vs. MeanFlow (Geng 2025): The core idea stems from the image generation version of MeanFlow; this work implements it for robot actions, adds Elastic Time Horizons, and language-conditioned CFG.
  • vs. ACT (Zhao 2023): ACT utilizes action chunking and heuristic temporal ensembling to reduce inference pressure but is prone to over-smoothing; ElasticFlow enforces physics consistency at the formula level and allows \(\Delta t\) to dynamically adjust chunk length.

Rating

  • Novelty: ⭐⭐⭐⭐ First application of MeanFlow in robot control combined with a clear Elastic Time Horizon mechanism.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Comprehensive simulation (LIBERO, CALVIN, RoboTwin, RoboCasa) + xArm6 real-robot tests + latency scans; Mismatch tests are cleverly designed.
  • Writing Quality: ⭐⭐⭐⭐ Clear derivations, intuitive analogies ("Spectral Zoom Lens"), and geometric explanations; dense but well-structured.
  • Value: ⭐⭐⭐⭐⭐ Provides a viable 1-NFE paradigm for VLA/robot diffusion policies with both engineering and methodological merit.