acceptodds
Under review as a conference paper at ICLR 2027

TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on a Consumer-Grade GPU with <1 GB VRAM

Abstract

Vision-language-action (VLA) models commonly adopt an LLM-centric pathway, processing visual observations and language instructions through a large language model before predicting robot actions. Although effective, this design incurs substantial computation and memory overhead. In this work, we introduce , a compact VLA architecture built on a direct mapping. Instead of using a large language model as the central interface between perception and action, TurboVLA independently encodes visual observations and language instructions, directly exchanges information between them through lightweight bidirectional vision-language interaction, and predicts continuous action chunks with a compact decoder. This simple design directly constructs task-conditioned representations while avoiding the overhead of an LLM-centered execution pathway. On LIBERO, TurboVLA achieves 97.6% average success with only 0.2B parameters, 31.2ms inference latency, and 0.9GB inference VRAM on a consumer-grade RTX 4090. Notably, a 0.4B TurboVLA achieves 88.06% success on RoboTwin 2.0, even matching or outperforming substantially larger VLA policies. These results demonstrate that the simple design of TurboVLA can achieve high performance without requiring an LLM-centric execution pathway, offering a new perspective on how vision, language, and action can be connected for efficient robotic manipulation. Code will be available.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-language-action models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_axLEOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

TurboVLA replaces the LLM-centered pathway in a vision-language-action model with separate visual and text encoders, lightweight bidirectional interaction, and an action-chunk decoder. The paper evaluates this design on LIBERO, RoboTwin 2.0, and four real-robot tasks, reporting competitive manipulation success with substantially lower model size, inference latency, and GPU memory use than the compared VLAs.

Critical strengths

TurboVLA achieves high success rates on both LIBERO and RoboTwin with a much smaller model and lower measured inference latency than the compared VLAs. This is a meaningful result for deploying manipulation policies with limited compute.

The paper tests one jointly trained policy across many tasks, includes a four-task real-robot comparison, and ablates language conditioning and vision–language interaction. These experiments show that the lightweight architecture can support more than a single memorized manipulation skill.

Critical weaknesses

The paper shows competitive success rates and strong inference efficiency on multi-task benchmarks, but its generalization tests largely cover trained task categories and predefined variations. It does not systematically evaluate unseen objects, scenes, camera viewpoints, or perturbations outside the training distribution. It therefore remains unclear whether removing the LLM trades away visual or language generalization for efficiency. This limits the scope of the claim that a lightweight architecture can replace an LLM-centric execution pathway for robot manipulation.

Questions for authors

Please report TurboVLA’s performance when trained only on clean RoboTwin 2.0 demonstrations and evaluated in the randomized setting. The current model is trained on both clean and randomized demonstrations, so its randomized score does not establish robustness to unseen perturbations.

Please evaluate the existing LIBERO-trained checkpoint on LIBERO-Plus without further training and report results by perturbation type. This would show whether its high success rate on standard LIBERO holds under distribution shifts.

Reviewer_csNjOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TurboVLA asks whether execution-level manipulation needs a large language model between perception and action. It encodes camera views with DINOv3 and the instruction with frozen BERT, fuses them with six bidirectional cross-attention layers initialized from Grounding DINO, and decodes action chunks with an ACT-style decoder trained by ℓ1 behavior cloning. Unlike π0-style VLAs, no language-model backbone is used; the contribution is the combination of existing components and the efficiency evidence. Validation covers LIBERO (0.2B, 97.6%, 31.2 ms, 0.9 GiB on an RTX 4090), RoboTwin 2.0 (0.4B, 88.06%), four real-robot tasks against π0.5, and ablations.

Critical strengths

The cost reduction is large and measured with a stated protocol (Appendix B.2). Against π0.5 on LIBERO, TurboVLA uses 0.2B instead of 3.4B parameters, 0.9 instead of 7.7 GiB and 31.2 instead of 238.8 ms at similar success (97.6 vs 96.9, Table 1); on RoboTwin 2.0 it is 5.7 times faster with higher success (Table 2). This directly supports the central claim and is the main reason for a positive rating.

The interaction ablation holds on the harder benchmark: removing interaction drops RoboTwin 2.0 from 88.06 to 82.58, with bidirectional interaction best under both the clean and randomized conditions (Table 10).

Critical weaknesses

The headline frequency is not exercised. 32 Hz is 1/31.2 ms of policy inference from GPU-resident inputs (Appendix B.2), yet Figure 1 says actions update at 32 Hz. Every chunk is executed in full before replanning (Section 3.1; Appendix B.1, H = 12 on LIBERO), so no result shows that faster inference helps control, and real-robot latency is not reported.

The success comparisons are imprecise. Each result comes from a single reported model with a fixed evaluation seed (Appendix B.1), and no variance across training runs is reported. Margins over the next-best method (0.2 points on LIBERO, 0.22 on RoboTwin) are below one binomial standard error (about 0.34 and 0.33). Several baseline success rates in Table 1 (e.g., the Diffusion Policy and OpenVLA rows) appear to match previously published values, while all latencies were re-measured (Appendix B.2); the paper does not state which success rates were reproduced. Grounding DINO initialization is not ablated. This weakens the outperforming wording more than the decision.

Questions for authors

New experiment: can you report results over at least three training seeds, including the no-interaction variant and randomly initialized interaction layers?

Clarification: which baseline success rates were reproduced, and what precision and denoising settings were used for baseline latency?

Clarification: on the real robot, how many actions per chunk were executed before replanning, and what control rate was achieved?

Reviewer_qnX9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TurboVLA investigates whether execution-level language-conditioned manipulation can achieve strong performance with a compact policy that does not rely on a large generative language model. It combines a DINOv3 visual encoder, a frozen BERT instruction encoder, a bidirectional vision–language interaction module initialized from Grounding DINO, and an ACT-style action-chunk decoder. The paper reports 97.6% success on LIBERO with 216.1M total parameters, 31.2 ms local policy inference latency, and 0.9 GiB inference memory on an RTX 4090. Evaluation on RoboTwin 2.0 and four physical manipulation tasks provides additional evidence, while ablations examine language conditioning, text encoders, interaction design, and action horizon. The main contribution is a practically useful compact-policy recipe and its reported performance–efficiency trade-off.

Critical strengths
  1. The reported combination of strong task performance and low inference cost is practically valuable. Achieving 97.6% success on LIBERO with a compact policy, 31.2 ms local inference latency, and a reported 0.9 GiB memory footprint provides useful evidence that execution-level manipulation can benefit from substantially lighter architectures. I view the contribution primarily as an effective integration and empirical design result, which can be meaningful even though the individual components have established precedents.

  2. The evaluation has useful breadth. Results span two simulation benchmarks and physical manipulation, with task-level appendix results and ablations covering several important design choices. The text-encoder and interaction ablations, including additional RoboTwin experiments, strengthen the evidence that the proposed recipe is effective beyond a single benchmark configuration. This breadth is a substantial reason for my positive assessment.

Critical weaknesses
  1. The central performance–efficiency comparison would benefit from clearer configuration traceability. Appendix B.2 provides a useful common timing protocol, but the main tables do not fully identify the provenance of each success score or establish its correspondence to the checkpoint, observation inputs, precision, and runtime settings used for efficiency measurements. Since the favorable quality–cost trade-off is central to the contribution, this correspondence matters. I regard this as a fixable reporting limitation rather than evidence that the comparisons are incorrect.

  2. Some interpretations are stronger than the measurements directly establish. In particular, 31.2 ms measures local action-chunk inference, whereas the LIBERO evaluation executes the full chunk before obtaining another observation. The approximately 32 Hz inference capacity should therefore be distinguished from a measured closed-loop observation-update rate. Similarly, the fusion ablations support the effectiveness of the complete configuration but do not fully separate interaction direction, capacity, and transferred pretrained knowledge. These limitations mainly call for more precise framing and do not undermine the usefulness of the overall policy recipe.

Questions for authors
  1. Clarification: For TurboVLA and the principal baselines in Tables 1–2, could you identify the success-score source, checkpoint, observation inputs, precision, and runtime settings, and confirm whether the success and efficiency results correspond to the same policy configuration? This would strengthen confidence in the central performance–efficiency comparison.

  2. Clarification: Could you distinguish local chunk-inference capacity from the observation and command schedule used during deployment, and align the title/Figure 1 terminology accordingly? For the 7.42 ms optimized implementation, please also clarify how changing inputs are refreshed and whether output equivalence or policy performance relative to the unoptimized implementation has been checked.

  3. Clarification: Which Grounding DINO submodules are transferred, and how do the interaction ablations differ in pretrained initialization and parameter count? An explicit description would help establish the reproducible recipe and the scope of the conclusions supported by these ablations. A separate experiment isolating each factor would only be necessary if a stronger mechanism-specific claim is intended.

Reviewer_7BVgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces TurboVLA, a lightweight vision-language-action model for execution-level robotic manipulation. It uses DINOv3 and BERT to encode visual observations and language instructions separately, fuses their features through bidirectional cross-attention, and predicts continuous action chunks with an ACT-style Transformer decoder trained using behavior cloning and an ℓ1\ell_1 loss. With 0.2B parameters, TurboVLA achieves 97.6% average success on LIBERO, with 31.2 ms inference latency and 0.9 GiB of inference memory; its 0.4B configuration achieves 88.06% average success on RoboTwin 2.0. The paper also reports real-world evaluations on an AgileX Piper robot. These results suggest that competitive execution-level manipulation does not necessarily require an LLM-centric pathway between perception and action.

Critical strengths
  1. The paper formalizes LLM-centric VLAs as a V→L→AV\to L\to A representational pathway. This turns the question of whether execution-level control needs an LLM at its center into a testable architectural hypothesis. On LIBERO, the method reaches 97.6% average success with 0.2B parameters, 0.9 GiB of inference memory, and 31.2 ms latency, an informative efficiency–performance trade-off.

  2. The efficiency evaluation protocol is clearly specified (Appendix B.2, Algorithm 1). The interaction-design ablation yields the same ranking on both LIBERO and RoboTwin 2.0. On RoboTwin 2.0, bidirectional interaction outperforms simple concatenation by 5.48 percentage points, which is fairly robust evidence.

Critical weaknesses
  1. The causal attribution of the central claim is not isolated. All supporting evidence compares the proposed method (~0.2B, no LLM backbone) with LLM-centric methods of 3.4–8.3B, so model scale and fusion architecture vary together. In addition, models of 2.8B–9.0B in Table 1 all fall within 96.4%–97.4%, which suggests the benchmark may be close to saturation. These results therefore do not establish that removing the LLM pathway itself is responsible for the observed performance.
  2. The role of the language branch is not validated. In Table 3, a learned Task-ID embedding trails semantic instructions by only 1.0 percentage point. Because LIBERO instructions map one-to-one onto tasks, this result cannot tell whether the model understands instruction semantics or uses instructions only as task identifiers.
  3. The scope of "execution-level" control is not operationally defined. It is illustrated only by two extreme examples, "stack the three bowls" and "clean the room," and the claim in the Limitations that open-ended commands require high-level reasoning and task decomposition has no experimental support.
Questions for authors
  1. Could the authors build a controlled baseline of comparable scale (~0.2B) that fuses visual and textual features within a single-sequence, LLM-style backbone, keeping the training data, training steps, action horizon, decoder, and loss unchanged, and report its performance relative to the proposed method?

  2. Given that the Task-ID embedding trails semantic instructions by only 1.0 percentage point, to what extent does the policy's behavior depend on instruction semantics rather than on identifying the task from visual observations? For example, does the policy's behavior change when task B's instruction is given under task A's observation? Is performance preserved under paraphrased instructions?

  3. The central claim is restricted to "execution-level" control, but this scope is currently defined only by two extreme examples. Could the authors define it more precisely, for example by stating the conditions under which the method applies in terms of instruction abstraction, number of subgoals, or task horizon, and provide supporting experimental evidence?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.