Driving · 2026-10-09

SimVLA: Zero-Shot Sim-to-Real VLA Learning for Mobile Manipulation

Kyoungin Baik, Youngwoon Lee

arXiv:2610.11248PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

SimVLA is an end-to-end, zero-shot sim-to-real VLA framework for mobile manipulation trained entirely on synthetic simulation data without teleoperation. It pre-trains on SimAction (35 tasks via skill composition) and SimVQA (privileged simulator-derived VQA supervision), then post-trains on a mixture of SimAction and SimDeploy (policy rollouts). The authors report zero-shot transfer to real environments and better performance than a π0.5 fine-tuned with 50 in-domain real demonstrations.

Why it matters

The authors target the data bottleneck of robot VLA learning for long-horizon mobile manipulation by using simulation for multiple supervision types beyond action generation. They report that zero-shot policies trained only in simulation can outperform fine-tuned baselines using 50 in-domain real demonstrations, and that the approach remains more robust in out-of-distribution real settings (pantry/real home).

Method

  • Pre-train SimVLA on two synthetic datasets: SimAction (action trajectories across 35 mobile manipulation tasks) and SimVQA (VQA-style visual-language supervision from privileged simulator state).
  • Construct SimDeploy by rolling out the SimVLA policy zero-shot across diverse simulated environments, collecting 22K trajectories (460 hours) to capture deployment-induced state distribution shift.
  • Train in two stages (pre-train on SimAction+SimVQA, then post-train on a mixture of SimAction and SimDeploy) to improve zero-shot sim-to-real transfer for long-horizon mobile manipulation.
Abstract (from arXiv)

Large-scale, diverse datasets have driven the success of LLMs and VLMs. But VLAs for robotics remain limited by the cost and complexity of real-world data collection. While simulation offers a scalable alternative, its potential for sim-to-real VLA learning in mobile manipulation remains largely underexplored. We introduce SimVLA, an end-to-end framework that trains VLAs entirely on synthetic simulation data without teleoperation for mobile manipulation. SimVLA is first pre-trained on two complementary simulation-derived datasets: SimAction, a large-scale robot action dataset spanning 35 diverse mobile manipulation tasks, generated by composing atomic skills, and SimVQA, which leverages privileged simulator state to provide spatial, geometric, and subtask-level visual-language supervision. We further post-train SimVLA on a mixture of SimAction and SimDeploy, a dataset collected from policy rollouts across diverse simulated environments. We evaluate SimVLA on tasks including restocking, pouring, and cleaning, and show zero-shot transfer to real-world mobile manipulation, including real home environments. SimVLA outperforms policies trained on 50 in-domain real-world demonstrations, suggesting that simulation can enable scalable sim-to-real mobile manipulation. We further demonstrate the value of multiple complementary forms of supervision for effectively leveraging simulation in VLA training.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-08 · Indexed by Pub-AI: 2026-10-09 UTC

preprint

SimVLA: Zero-Shot Sim-to-Real VLA Learning for Mobile Manipulation

Kyoungin Baik, Youngwoon Lee

Preprint posted: 2026-10-08 · Original record ↗

arxiv
2610.11248
Explicit source versions (1)
  • v1 · 2026-10-08
  1. Preprint available

    Source date: 2026-10-08 · First detected: 2026-10-25 UTC · Historical baseline

    Source evidence (1)
Abstract · arxiv

Large-scale, diverse datasets have driven the success of LLMs and VLMs. But VLAs for robotics remain limited by the cost and complexity of real-world data collection. While simulation offers a scalable alternative, its potential for sim-to-real VLA learning in mobile manipulation remains largely underexplored. We introduce SimVLA, an end-to-end framework that trains VLAs entirely on synthetic simulation data without teleoperation for mobile manipulation. SimVLA is first pre-trained on two complementary simulation-derived datasets: SimAction, a large-scale robot action dataset spanning 35 diverse mobile manipulation tasks, generated by composing atomic skills, and SimVQA, which leverages privileged simulator state to provide spatial, geometric, and subtask-level visual-language supervision. We further post-train SimVLA on a mixture of SimAction and SimDeploy, a dataset collected from policy rollouts across diverse simulated environments. We evaluate SimVLA on tasks including restocking, pouring, and cleaning, and show zero-shot transfer to real-world mobile manipulation, including real home environments. SimVLA outperforms policies trained on 50 in-domain real-world demonstrations, suggesting that simulation can enable scalable sim-to-real mobile manipulation. We further demonstrate the value of multiple complementary forms of supervision for effectively leveraging simulation in VLA training.