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

AI summary

Overview

Research area: Robotics — bimanual (two-arm) robot manipulation with vision-language-action (VLA) models, cross-embodiment transfer, and modular policy composition.

Technical level: Advanced. The paper assumes familiarity with transformer architectures, mixture-of-experts layers, flow matching for action generation, and standard VLA training pipelines.

Scope: The paper proposes TwinVLA, a modular architecture that composes two copies of a pretrained single-arm VLA into one coordinated bimanual policy, and evaluates it against monolithic baselines on real-world and simulated two-arm tasks.

What This Paper Is About

Public robot datasets overwhelmingly contain single-arm demonstrations, while bimanual (two-arm) data is scarce. Training a strong two-armed robot policy has therefore required either collecting thousands of hours of proprietary bimanual data or pretraining a single large model on a mixture of one-arm and two-arm data. The authors ask whether a bimanual policy can instead be built by composing two copies of an already-pretrained single-arm VLA and teaching them to coordinate using only a small number of bimanual demonstrations.

Key Contributions

1. A modular bimanual architecture. TwinVLA duplicates a pretrained single-arm VLA into left and right branches and links them through joint attention (shared self-attention layers) plus a mixture-of-experts (MoE) module for shared inputs.

2. A data-efficient adaptation paradigm. The method requires no bimanual pretraining at all. It is fine-tuned on only 50 episodes of target bimanual data for real-world tasks, after pretraining on approximately 800 hours of public single-arm OXE data.

3. Strong empirical results against far larger-data baselines. Across real-world and simulated bimanual tasks, TwinVLA outperforms the comparably sized monolithic RDT-1B (1.2B parameters) and approaches the 3.3B-parameter π₀, which uses extensive proprietary bimanual data and compute.

4. A systematic ablation of the coordination mechanisms. The paper isolates the contribution of attention re-weighting, MoE integration, joint attention, single-arm pretraining, and the twin structure itself.

Main Findings

• Real-world performance beats the comparable monolithic baseline. On five long-horizon real-world bimanual tasks (Fold towel, Extract hexkey, Carrot to bag, Brush to dustpan, Take towel off, plus the Put X into pot task set) on an Anubis dual-arm robot, TwinVLA outperforms RDT-1B on average. Diffusion Policy (DP, 271M parameters) performs poorly, confirming the importance of pretraining; π₀ achieved the highest overall performance at significantly higher cost.

• TwinVLA outperforms RDT-1B by 26.0% in the real world, 5.0% in simulation, and 21.8% in language-following tasks on average.

• Simulation results are competitive across a large task suite. On RoboTwin 2.0 (50 bimanual tasks, 50 generated demonstrations per task, 100 test rollouts under Easy and Hard settings), TwinVLA beats RDT-1B in the Easy setting by 7.48%. On the Hard setting it is 3.72% lower than RDT-1B but still outperforms a non-pretrained Diffusion Policy by 9.38%. On Tabletop-Sim (5 tasks: dish-drainer, handover-box, shoes-table, lift-box, and put X box into Y pot; 50 episodes per task, 500 evaluation rollouts), TwinVLA outperforms RDT-1B in most scenarios, except the RoboTwin Hard tasks, and in Tabletop-Sim Easy tasks it even outperforms π₀.

• Robustness. In the Tabletop-Sim Hard setup (varied textures and objects), TwinVLA outperforms RDT-1B by 3.3%. On the real-world Fold towel task under low light, TwinVLA reached 45.0% success versus π₀'s 40.0% and RDT-1B's 15.0%; with distractors present, TwinVLA reached 25.0% versus π₀'s 60.0% and RDT-1B's 15.0% — the authors note TwinVLA is robust to lighting changes but less effective with distractors.

• Data efficiency during fine-tuning. On Tabletop-Sim Easy tasks, TwinVLA's success rate rises steeply from 20 to 35 to 50 demonstrations and surpasses RDT-1B at 50 demonstrations.

• Language following is preserved. On the "Put X into pot" task in both simulation and real-world settings, TwinVLA outperforms both RDT-1B and π₀.

• Joint attention is the most important coordination mechanism. Sequentially removing components from the full model caused performance drops of 1.1% (simulation) and 4.0% (real world) for attention re-weighting; an additional 1.1% and 5.0% for MoE; and the largest additional drops of 4.0% and 27.0% for joint attention. Removing attention re-weighting increased initial fine-tuning loss by 40%.

• Removing MoE costs efficiency. Without MoE, token sequence length increased by 28% and VRAM usage increased by 21%.

• Single-arm pretraining is essential. Training from scratch without the OXE-based pretraining caused a 4.6% drop in simulation and a 46.0% drop in the real world.

• Resource footprint. TwinVLA has 1.3B parameters versus RDT-1B's 1.2B and π₀'s 3.3B. TwinVLA uses approximately 800 hours of single-arm data and 50 episodes of target bimanual data plus approximately 25 H100 GPU-days, compared to approximately 2,400 hours and approximately 1,440 H100 GPU-days for RDT-1B, and approximately 10,900 hours with more than 1,000 H100 GPU-days for π₀. The paper states RDT-1B reportedly took a month on 48 H100 GPUs, and π₀ relies on a 10,000-hour proprietary dataset.

Methodology in Plain English

The authors first build a small single-arm VLA called SingleVLA, based on the Eagle2-1B vision-language model but with the language head removed, bringing it to 0.8B parameters. They pretrain it on a roughly 800-hour subset of the OXE (Open X-Embodiment) dataset for 120k steps, about 5 days on 5 H100 GPUs. All actions across all source robots are converted into a shared representation: absolute end-effector poses with a 6D rotation encoding, relative to the robot's base frame, producing a 10-dimensional action space. To handle datasets recorded at different control rates, all data is resampled to 20 Hz via interpolation (frequency matching), avoiding misalignment when an action chunk of 20 steps means about 7 seconds in RT-1 (3 Hz) but about 1.3 seconds in DROID (15 Hz).

To form TwinVLA, they copy this pretrained model for the left and right arms — but not entirely. The vision encoder and the DiT action head are shared, while the VLM backbone is fully duplicated, and each arm gets its own small proprioception encoder. This yields a 1.3B-parameter model.

The two duplicated VLMs are tied together by joint attention: at each transformer layer, the query, key, and value tensors from both arms are concatenated, self-attention is applied over the union, and the outputs are split back into the two streams. Everything else (projections, feed-forward networks) stays arm-specific. A custom causal attention mask keeps generation autoregressive: lower-triangular masking within each arm's region, the shared modality fully visible, and each arm allowed to attend to half of the other arm's tokens.

Because the shared inputs (language instruction and ego-centric image) would otherwise be fed redundantly to both models, a mixture-of-experts layer routes shared tokens between the two VLMs' feed-forward networks using softmax weighting, and projections and layer norms use an output-averaging strategy rather than physically merged parameters. This reduces VRAM by 21%, allowing a batch size of 8 on a single 40 GB GPU. Lastly, attention re-weighting rescales attention scores for the shared modality so new arm-specific tokens don't disrupt pretrained attention patterns.

Actions are generated with conditional flow matching: an action head predicts a flow from a noised action chunk to the target, sampled by forward Euler integration with 10 steps. TwinVLA fine-tuning used 100k steps with a global batch size of 8, learning rate 1×10⁻⁴ with cosine schedule, warm-up ratio 0.05, AdamW (weight decay 1×10⁻⁵, epsilon 1×10⁻⁸), frozen vision backbone, no image augmentation, and action chunk size 20 — about 2 days on a single L40S GPU.

Why This Matters

Impact on research. The paper reframes bimanual manipulation as a coordination problem rather than a data-collection problem. It shows that modular composition of pretrained single-arm policies can substitute for expensive bimanual pretraining, and that the inductive bias of a twin structure beats a monolithic model of comparable size. It also offers a reproducible alternative to proprietary-dataset approaches, since it relies only on public single-arm data. The appendices additionally validate SingleVLA's VLM choice on the LIBERO benchmark using several VLMs and 500 rollouts per task suite, though the specific comparison table is truncated in the supplied content.

Real-world applications:

• Household robots performing coordinated two-handed chores such as folding towels, moving dishes to a drainer, or wiping surfaces.
• Warehouse and logistics robots that need two arms to lift, hand off, or pack objects — TwinVLA was evaluated on tasks including handover-box and lift-box.
• Industrial assembly and tool-handling tasks, such as extracting a hexkey or manipulating parts with two arms.
• Service and assistive robotics operating under imperfect lighting or clutter, given the tested robustness to low-light conditions.

Industry relevance. The economic argument is direct: TwinVLA is reported to reach comparable or better performance with roughly 25 H100 GPU-days versus more than 1,000 for the baselines, and with about 800 hours of public single-arm data versus thousands of hours of proprietary bimanual data. For companies without proprietary bimanual datasets, this lowers the barrier to building capable two-arm policies. The approach also preserves the language-following and scene-robustness properties that make VLAs attractive in products.

Future Directions

• Resolving the visual disparity between arms. The authors state that generalization remains limited because the two-arm visual configuration differs from the single-arm pretraining distribution; mechanisms to prevent this could help integrate diverse data and improve generalization to unseen tasks.

• Alternative action representations. The paper uses absolute end-effector pose control for its embodiment-agnostic properties; the authors suggest exploring relative absolute actions or shared representations to further improve transfer efficiency.

• Beyond bimanual manipulation. The authors position TwinVLA as a blueprint for other dataset-imbalanced domains, explicitly mentioning mobile manipulation as a promising extension.

• Improved explainability. The limitations section also lists improving model explainability as an open goal alongside better generalization.

Target Audience

Robotics and embodied-AI researchers working on manipulation policies and VLA architectures; engineers building two-armed robot systems who need to work with limited in-house bimanual data; and ML practitioners interested in modular composition, mixture-of-experts routing, and cross-embodiment transfer as alternatives to scaling monolithic models. Readers should already be comfortable with transformer internals, attention masking, and flow-matching action heads, since the core contributions are architectural.

Authors’ abstract

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 a comparably-sized monolithic RDT-1B model without requiring any bimanual pretraining. Furthermore, it narrows the gap to state-of-the-art model $π_0$, which relies 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 public single-arm data.