Compose two pretrained single-arm VLAs into one bimanual policy — data-efficient, high-performance bimanual manipulation without any bimanual pretraining.
Built only from public single-arm data: two copies of a single-arm VLA, linked by Joint Attention.
Outperforms the comparable-size monolithic model with less pretraining data (0.5M vs 1.4M) and compute.
Narrows the gap to π0, which relies on extensive proprietary bimanual data.
Vision-language-action models (VLAs) trained on large-scale robotic datasets have demonstrated strong performance on manipulation tasks, including bimanual tasks. However, because most public datasets focus on single-arm demonstrations, adapting VLAs for bimanual tasks typically requires substantial additional bimanual data and fine-tuning. To address this challenge, we introduce TwinVLA, a modular framework that composes two copies of a pretrained single-arm VLA into a coordinated bimanual VLA. Unlike monolithic cross-embodiment models trained on mixtures of single-arm and bimanual data, TwinVLA improves both data efficiency and performance by composing pretrained single-arm policies. Across diverse bimanual tasks in real-world and simulation settings, TwinVLA outperforms its direct competitor RDT-1B, a monolithic model of comparable size without requiring any bimanual pretraining. Furthermore, it narrows the gap to state-of-the-art model, π0 which rely on extensive proprietary bimanual data and compute cost. These results establish our modular composition approach as a data-efficient and scalable path toward high-performance bimanual manipulation, leveraging existing public single-arm data.
TwinVLA duplicates a VLM backbone pretrained on single-arm data (Left) into two arm-specific branches linked by Joint Attention (Right). Shared inputs (ego-centric views, language) go through a mixture-of-experts, and only the backbone is duplicated, so the model stays small.
SingleVLA is a lightweight single-arm VLA pretrained on large-scale data. All actions are standardized to absolute end-effector poses with a 6D rotation, which keeps control consistent across robots and makes the extension to two arms direct.
The two backbones share only their self-attention layers. A causal mask lets each arm see the shared inputs and the other arm's tokens symmetrically, without breaking the autoregressive order.
Shared inputs are processed once through a MoE, which shortens the token sequence and saves VRAM. Training is stabilized with a simple attention re-weighting.
We evaluate on bimanual tasks in simulation and the real world against RDT-1B (a monolithic model of comparable size), π0 (a larger model trained on extensive proprietary bimanual data), and Diffusion Policy (no pretraining).
On three coordinated tabletop tasks on the dual-arm Anubis robot, TwinVLA outperforms RDT-1B while pretrained on far less single-arm data (0.5M vs 1.4M). DP trails, showing the value of pretraining; π0 performs best overall.
On RoboTwin 2.0 and our Tabletop-Sim, TwinVLA outperforms RDT-1B in most settings and is comparable to π0, even surpassing it on the Tabletop-Sim Easy tasks.
On multi-task Tabletop-Sim with nuanced instructions, TwinVLA outperforms both RDT-1B and π0: fine-tuning keeps the VLM's instruction following.
TwinVLA adapts quickly to new bimanual tasks and surpasses RDT-1B with just 50 demonstrations.
Removing Joint Attention, MoE, or attention re-weighting each hurts performance, with Joint Attention mattering most. Single-arm pretraining is essential, and the twin structure beats a comparable monolithic model.
@inproceedings{
im2026twinvla,
title={Twin{VLA}: Data-Efficient Bimanual Manipulation with Twin Single-Arm Vision-Language-Action Models},
author={Hokyun Im and Euijin Jeong and Andrey Kolobov and Jianlong Fu and Youngwoon Lee},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=jG9W6nAwVz}
}