Driving · 2026-10-05
FastOPD: On-Policy Distillation for Lightweight VLA Deployment
Yoojin Oh, Jeongsol Kim, Yeonwoo Seo, Jangho Park, Seonghyun Jin, Sunwoo Park, Youngmin Kim, Youngjun Jun, Kyumin Choi, Jong Chul Ye
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
FastOPD is an on-policy distillation framework to deploy lightweight Vision-Language-Action (VLA) models by transferring dynamics from a large flow-map teacher using only single-state teacher supervision plus a self-consistency objective. The authors claim theory that minimizing the objective recovers the few-step distribution induced by the teacher’s ideal shortcut model. They evaluate in simulation and on a real robot.
Why it matters
The paper targets a deployment bottleneck of large VLA/flow-based policies: high training/inference cost. By reducing teacher evaluations during on-policy distillation and enabling few-step inference, FastOPD aims to keep performance while making distilled students practical for real-world robots.
Method
- Adapts a flow map for single-state teacher supervision: the student is matched to the teacher only at sampled on-policy states rather than every denoising step.
- Combines on-policy flow-map distillation (velocity matching at one sampled state) with a self-consistency objective that propagates local velocity supervision to finite-interval shortcut transitions via a midpoint.
Abstract (from arXiv)
Vision-Language-Action (VLA) foundation models have scaled rapidly to enhance manipulation performance and generalizability, but this scaling incurs high computational costs that render real-world deployment increasingly challenging. Existing approaches typically mitigate this issue by designing smaller architectures or reducing the iterative denoising steps in flow-based policies. In this work, we propose FastOPD, a foundation-to-lightweight VLA framework that enables the practical deployment of large-scale VLAs through efficient on-policy distillation. Specifically, FastOPD adapts a flow map for single-state teacher supervision and combines it with a self-consistency objective to construct a compact student that learns the teacher dynamics. Furthermore, we theoretically demonstrate that minimizing this objective allows the distilled student to recover a distribution on par with that induced by an ideal few-step teacher model. We evaluate FastOPD across diverse foundation policies in simulation and real-world experiments. On LIBERO, FastOPD retains 84% of the performance of $\pi_{0.5}$ with only two inference steps, reducing inference latency by 78.1% while outperforming existing few-step distillation baselines in average success rate. With LingBot-VLA as the teacher, FastOPD improves the single-step success rate over the base student by 15.9 percentage points on RoboTwin 2.0. We further demonstrate its applicability to a World Action Model (WAM) and deploy a compact student distilled from MolmoAct2 on a real robot.
Related papers
- DriftOPD: Sequence-Level Reverse-KL Distillation for One-Step VLA Policies
- VeriVLA: Adaptive Test-Time Verification for Flow-Based VLAs with Latent World Models
- Fast-VLA: You Only Think Once for Robot Manipulation
- WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction
- MIKASA-Robo-VLA: Benchmarking Memory in VLA Models for Long-Horizon Manipulation
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-02 · Indexed by Pub-AI: 2026-10-05 UTC
FastOPD: On-Policy Distillation for Lightweight VLA Deployment
Preprint posted: 2026-10-02 · Original record ↗
- arxiv
- 2610.02832
Explicit source versions (1)
- v1 · 2026-10-02
- Preprint available
Source date: 2026-10-02 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Vision-Language-Action (VLA) foundation models have scaled rapidly to enhance manipulation performance and generalizability, but this scaling incurs high computational costs that render real-world deployment increasingly challenging. Existing approaches typically mitigate this issue by designing smaller architectures or reducing the iterative denoising steps in flow-based policies. In this work, we propose FastOPD, a foundation-to-lightweight VLA framework that enables the practical deployment of large-scale VLAs through efficient on-policy distillation. Specifically, FastOPD adapts a flow map for single-state teacher supervision and combines it with a self-consistency objective to construct a compact student that learns the teacher dynamics. Furthermore, we theoretically demonstrate that minimizing this objective allows the distilled student to recover a distribution on par with that induced by an ideal few-step teacher model. We evaluate FastOPD across diverse foundation policies in simulation and real-world experiments. On LIBERO, FastOPD retains 84% of the performance of $π_{0.5}$ with only two inference steps, reducing inference latency by 78.1% while outperforming existing few-step distillation baselines in average success rate. With LingBot-VLA as the teacher, FastOPD improves the single-step success rate over the base student by 15.9 percentage points on RoboTwin 2.0. We further demonstrate its applicability to a World Action Model (WAM) and deploy a compact student distilled from MolmoAct2 on a real robot.