TwinVLA: Data-Efficient Bimanual Manipulation with Twin Single-Arm Vision-Language-Action Models

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; TwinVLA: data-efficient bimanual manipulation via VLAs

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

TwinVLA composes two copies of a pretrained single-arm Vision-Language-Action (VLA) model into a coordinated bimanual policy via joint attention and Mixture-of-Experts, eliminating the need for large-scale bimanual pretraining data. It achieves competitive bimanual manipulation performance using only ~800h of public single-arm data and 50 fine-tuning episodes, versus ~2,400h (RDT-1B) and ~10,900h (π0). The approach is both data- and compute-efficient, requiring just 25 H100 GPU-days.

Problem

Most public robotic datasets are single-arm, so adapting VLAs to bimanual tasks via the prevailing monolithic cross-embodiment approach (e.g., RDT-1B, π0) demands thousands of hours of proprietary bimanual data and massive compute (>1,000 H100 GPU-days), limiting reproducibility. Existing modular designs impose backbone-specific constraints that prevent integration with general-purpose VLA frameworks.

Method

TwinVLA instantiates two copies ("twins") of a pretrained single-arm model (SingleVLA, 0.8B parameters, pretrained on the Open X-Embodiment dataset) and links them with three mechanisms:

1. Selective duplication: Only the VLM backbone is duplicated per arm; the vision encoder and DiT action head are shared, yielding a 1.3B-parameter total.
2. Joint attention for cross-arm fusion: At each transformer layer, queries, keys, and values from both VLM streams are concatenated and a single self-attention operation is performed, then split back. A causal attention mask embeds per-arm lower-triangular masks while allowing each arm to attend to half of the other arm's tokens. This is inspired by the Mixture-of-Transformers (MoT) mechanism.
3. MoE for shared inputs: Shared tokens (egocentric image, language instruction) are routed through both VLM backbones via a 2-expert MoE: \(\text{MoE}(x) = w_\text{left} \cdot \text{FFN}_\text{left}(x) + (1-w_\text{left}) \cdot \text{FFN}_\text{right}(x)\), reducing VRAM by 21%. Non-FFN shared layers use output averaging (task arithmetic–style).

Attention re-weighting rescales attention scores for the shared modality to prevent new arm-specific tokens from diluting pretrained attention patterns. Training uses conditional flow matching (Equation 1) with a DiT action head. Fine-tuning on bimanual tasks requires only ~50 demonstrations per task.

Key Contributions

• Modular dual-VLM architecture linking twin single-arm VLAs via joint attention with causal cross-arm masking, avoiding bimanual pretraining entirely.
• Data-efficient adaptation paradigm: ~800h single-arm pretraining data + 50 bimanual episodes per task, vs. 2,400–10,900h for comparable baselines.
• Compute-efficient training: 25 H100 GPU-days vs. >1,000 for RDT-1B and π0.
• Attention re-weighting to preserve pretrained modality importance during fine-tuning.
• Tabletop-Sim benchmark: 5 dexterous bimanual tasks in dm_control with easy/hard splits.

Results

• Real-world (5 tasks, Anubis robot, 20 rollouts each): TwinVLA avg. 65% vs. RDT-1B avg. 45% vs. Diffusion Policy avg. 20%; π0 (upper bound) avg. 71%.
• RoboTwin 2.0 (50 tasks, 100 rollouts):
• Easy: TwinVLA 46.4%, RDT-1B 42.0%, DP 13.7%; π0 72.5%
• Hard: TwinVLA 34.5%, RDT-1B 39.8%, DP 16.4%; π0 61.6%
• Tabletop-Sim (5 tasks, 500 rollouts):
• Easy: TwinVLA 45.6%, RDT-1B 37.4%, DP 21.4%; π0 42.4% (TwinVLA surpasses π0 here)
• Hard: TwinVLA 45.2%, RDT-1B 42.9%, DP 18.7%; π0 (not reported separately)
• Data efficiency (Tabletop-Sim Easy): TwinVLA with 50 demos surpasses RDT-1B; with 20 demos already competitive.
• Robustness (Fold towel, unseen conditions): Low-light: TwinVLA 45%, π0 40%, RDT 15%; Distractor: TwinVLA 25%, π0 60%, RDT 15%.

Limitations

• Underperforms RDT-1B on RoboTwin Hard tasks (34.5% vs. 39.8%), suggesting weaker generalization under large distribution shift without bimanual pretraining.
• Weaker robustness to visual distractors compared to π0, which benefits from extensive proprietary data diversity.
• SingleVLA must be pretrained first; the approach depends on availability of a suitable single-arm pretrained model.
• The method is validated on parallel-jaw grippers only; generalization to dexterous hands or other end-effectors is untested.
• Language-following evaluation results were truncated in the provided text.

Relevance to Vision-Language Models

TwinVLA demonstrates a practical and data-efficient strategy for scaling VLMs into multi-effector robotic control by composing pretrained models rather than retraining monolithic ones, directly relevant to modular VLA design. The joint attention mechanism adapts Mixture-of-Transformers—a technique from multimodal VLM literature—to synchronize two VLM backbones for physical coordination, showing cross-pollination between vision-language and embodied AI architectures. For VLM researchers, the MoE-based shared-token routing and attention re-weighting are generalizable techniques for composing pretrained transformer models with minimal interference to learned representations. The work also highlights a critical data bottleneck in embodied VLMs: the scarcity of bimanual robot data parallels broader challenges of domain-specific data scarcity in specialized VLM applications.