Unified Condition-Action Modeling for Accurate One-Step Action Generation

Xinyu Zhou; Zikun Cai; Kuangji Zuo; Gen Li; Boyu Ma; Yanshuo Lu; Yutong Song; Mingqi Yuan

cs.ROAug 17, 2026

Unified Condition-Action Modeling for Accurate One-Step Action Generation

Authors: Xinyu Zhou, Zikun Cai, Kuangji Zuo, Gen Li, Boyu Ma, Yanshuo Lu, Yutong Song, Mingqi Yuan,

Organizations: MARS Lab, Nanyang Technological University · The University of Hong Kong

Abstract

Robot manipulation requires policies that are both accurate and efficient, as robot control must respond to changing observations under tight latency constraints. Recent diffusion and flow policies are promising, but they often treat conditions as auxiliary signals rather than jointly evolving them with action trajectories. We find that this limitation can be effectively mitigated by a simple yet effective unified condition-action modeling design that represents conditions and actions in a shared token space, allowing a compact model to achieve high performance while improving both inference speed and accuracy. Therefore, we propose UCA-Flow, a unified condition-action modeling framework for accurate one-step action generation. Our method unifies observation conditions, timestep conditions, interval conditions, and action tokens into a single sequence, and processes them with a Unified Condition-Action Transformer for joint condition-action representation learning. As a result, condition representations are dynamically reconstructed according to the current generation stage, highlighting information most relevant for action refinement. Furthermore, we introduce an improved dual-pass supervision scheme over $u$ and $v$ for stronger optimization of unified condition-action modeling. UCA-Flow achieves an average success rate of 86.4%, improving over the strongest baseline by 7.5 percentage points, with a task-averaged inference latency of $3.0$ ms. It achieves $43.9\times$ and $32.2\times$ speedups over DP3 and Simple DP3, and remains $4.2\times$ and $2.2\times$ faster than one-step FlowPolicy and MP1, respectively. Project page: https://uca-policy.github.io/UCA.github.io/

Figures & tables

Figure 1: Motivation of UCA-Flow . (a) Existing policies often treat conditions as auxiliary signals. Unified condition-action modeling enables conditions and actions to evolve together, allowing condition representations to adapt to the current generation stage and action. (b) UCA-Flow achieves $86.4\%$ average success and $3.0$ ms task-averaged inference latency on 37 Adroit and Meta-World tasks, improving success over DP3 by 17.7 percentage points with a $43.9\times$ speedup. Speedups use task-weighted latencies before rounding.

Figure 2: Overview of UCA-Flow . Observation, timestep, interval, and noisy action are unified as condition-action tokens and processed by a Unified Condition-Action Transformer to predict the interval-aware velocity.

Figure 3: Gradient-based saliency visualization on real-world tasks . UCA-Flow highlights more task-relevant regions, including the end-effector, target object, and upper arm joints, compared with the auxiliary condition-injection AdaLN and Cross-Attention.

Method	Adroit		MetaWorld
Method	Door	Pen	Dial Turn	Peg Insert Side	Lever Pull	Pick Place Wall	Stick Pull	Shelf Place
UCA-Flow	$\mathbf{77.3\pm 2.1}$	$\mathbf{63.2\pm 3.4}$	$\mathbf{92.3\pm 4.7}$	$\mathbf{90.7\pm 4.0}$	$\mathbf{89.3\pm 3.2}$	$\mathbf{90.7\pm 3.2}$	$\mathbf{75.3\pm 2.3}$	$\mathbf{74.0\pm 3.5}$
-DualPass	$74.3\pm 5.0$	$57.0\pm 3.6$	$91.3\pm 7.3$	$88.0\pm 3.6$	$89.0\pm 0.0$	$89.6\pm 5.8$	$70.0\pm 3.5$	$64.0\pm 7.8$
-Unified/c	$68.2\pm 1.8$	$58.2\pm 6.9$	$87.3\pm 10.5$	$86.3\pm 3.2$	$79.0\pm 1.0$	$84.0\pm 3.4$	$62.3\pm 3.2$	$56.6\pm 3.8$
-Unified/a	$56.5\pm 2.1$	$48.6\pm 4.6$	$78.6\pm 9.3$	$61.0\pm 1.7$	$63.6\pm 6.0$	$60.3\pm 11.2$	$68.0\pm 3.5$	$32.3\pm 1.5$

Table 1: Ablation study on UCA-Flow components. -DualPass removes dual-pass supervision; -Unified/c and Unified/a replace unified condition-action modeling with cross-attention and AdaLN, respectively. Results are success rates on representative Adroit and Meta-World tasks.

Figure 4: Representative simulated rollouts. Rows correspond to Assembly, Adroit Door, and Drawer Close, and columns show successive stages of each execution. The trajectories require different interaction patterns: aligning and inserting an object, coordinating a dexterous hand with a door handle, and maintaining contact while pushing a drawer closed. UCA-Flow completes all three sequences from approach to task completion using one-step action generation.

Methods	Publication	NFE	Adroit			MetaWorld				Average
Methods	Publication	NFE	Hammer	Door	Pen	Easy (21)	Medium (4)	Hard (4)	Very Hard (5)	Average
DP3	RSS’24	10	$100\pm 0$	$56\pm 5$	$46\pm 10$	$87.3\pm 2.2$	$44.5\pm 8.7$	$32.7\pm 7.7$	$39.4\pm 9.0$	$68.7\pm 4.7$
Simple DP3	RSS’24	10	$98\pm 2$	$40\pm 17$	$36\pm 4$	$86.8\pm 2.3$	$42.0\pm 6.5$	$38.7\pm 7.5$	$35.0\pm 11.6$	$67.4\pm 5.0$
FlowPolicy	AAAI’25	1	$98\pm 1$	$61\pm 2$	$54\pm 4$	$84.8\pm 2.2$	$58.2\pm 7.9$	$40.2\pm 4.5$	$52.2\pm 5.0$	$71.6\pm 3.5$
MP1	AAAI’26	1	$100\pm 0$	$69\pm 2$	$58\pm 5$	$88.2\pm 1.1$	$68.0\pm 3.1$	$58.1\pm 5.0$	$67.2\pm 2.7$	$78.9\pm 2.1$
UCA-Flow	Ours	1	$\bm{100\pm 0}$	$\bm{77.3\pm 2.1}$	$\bm{63.2\pm 3.4}$	$\bm{92.4\pm 0.7}$	$\bm{79.3\pm 1.8}$	$\bm{68.0\pm 1.9}$	$\bm{85.3\pm 0.7}$	$\bm{86.4\pm 0.3}$

Table 2: Success rate comparison on 37 Adroit and Meta-World tasks . NFE denotes the number of function evaluations required for action generation. UCA-Flow group and overall statistics average tasks within each of three seeds before computing the mean and sample standard deviation across seeds; the overall mean weights all 37 tasks equally. Baseline summaries are retained as reported. The best results are highlighted in bold.

Figure 5: Training success curves on Adroit and Meta-World. Solid lines show success rates and shaded regions indicate variation across seeds. Using the top-five-checkpoint mean per seed, UCA-Flow outperforms the baselines on all four tasks, including an approximately 5% gain on Adroit Pen.

Figure 6: Real-world execution sequences. The top row shows Pick and Place progressing from the initial scene through grasping and transport to placement in the target box. The bottom row shows Drawer Close progressing from approach to contact, sustained pushing, and full closure. These sequences illustrate that UCA-Flow maintains task-relevant control through distinct free-space and contact-rich phases despite sensing and execution uncertainty.

Method	Pick & Place	Drawer Close
DP3	50.0	95.0
MP1	55.0	95.0
UCA-Flow	65.0	100.0

Table 4: Real-world comparison with policy baselines. Success rates (%) are measured over 20 independent trials per task. UCA-Flow improves Pick and Place by 10 percentage points over MP1 and achieves 100% success on Drawer Close.

Method	Pick & Place	Drawer Close
AdaLN	60.0	90.0
Cross Attention	50.0	95.0
UCA-Flow	65.0	100.0

Table 5: Real-world comparison of condition-integration designs. Success rates (%) are measured over 20 independent trials per task. Unified condition-action modeling performs best on both tasks, exceeding AdaLN by 5 and 10 percentage points on Pick and Place and Drawer Close, respectively.

One-Step Generative Modeling

One-step generative modeling maps noise or conditions to data in one network evaluation, replacing iterative denoising or numerical integration.

See all 147 papers

One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies
Mar 12, 2026Shaolong Li, Lichao Sun, Yongchao ChenVisuomotor Policy Learning One-Step Generative Modeling
CSE Department, University of Michigan · Lehigh University · College of AI, Tsinghua University
$Δ$ Flow Matching for Robot Foundation Models
Oct 11, 2026Kuangji Zuo, Tuo An, Xinying Guo +9Flow Matching for Action Generation Robot Foundation Models
MARS Lab, Nanyang Technological University
UniAction: Unified Action Conditioning for World Models
Oct 23, 2026Bowen Xue, Zizhang Li, Hong-Xing Yu +1Action-Conditioned World Models Action-Conditioned Video Generation
Stanford University

Mar 12, 2026J/K move · Enter open · S save

One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies

Shaolong Li, Lichao Sun, Yongchao Chen

CSE Department, University of Michigan · Lehigh University · College of AI, Tsinghua University

Visuomotor Policy Learning One-Step Generative Modeling Flow-Based Generative Modeling Robotic Manipulation Self-Distillation

Open paper AskPDF arXiv

Generative flow and diffusion models provide the continuous, multimodal action distributions needed for high-precision robotic policies. However, their reliance on iterative sampling introduces severe inference latency, degrading control frequency and harming performance in time-sensitive manipulation. To address this problem, we propose the One-Step Flow Policy (OFP), a from-scratch self-distillation framework for high-fidelity, single-step action generation without a pre-trained teacher. OFP unifies a self-consistency loss to enforce coherent transport across time intervals, and a self-guided regularization to sharpen predictions toward high-density expert modes. In addition, a warm-start mechanism leverages temporal action correlations to minimize the generative transport distance. Evaluations across 56 diverse simulated manipulation tasks demonstrate that a one-step OFP achieves state-of-the-art results, outperforming 100-step diffusion and flow policies while accelerating action generation by over

100\times

. We further integrate OFP into the

π_{0.5}

model on RoboTwin 2.0, where one-step OFP surpasses the original 10-step policy. These results establish OFP as a practical, scalable solution for highly accurate and low-latency robot control.

Abstract

One-Step Generative Modeling

Explore similar work

One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies

ΔΔΔ Flow Matching for Robot Foundation Models

UniAction: Unified Action Conditioning for World Models

$Δ$ Flow Matching for Robot Foundation Models