Pith.

Robotics · arXiv:2607.12659 v7 · submitted 14 Jul 2026

EagleVLA: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inference

Zebin Yang, Qi Wang, Yunhe Wang, Xiurui Guo, Bo Yu, Shaoshan Liu, Jiafeng Xu, Hao Dong

show 1 more author

Meng Li

The paper says

“Vision-Language-Action (VLA) models have achieved impressive performance on diverse embodied tasks.”

more of the abstract

However, deploying VLA models on low-power onboard devices, such as the Jetson Orin, remains challenging due to their high computational complexity, which leads to substantial inference latency and low control frequency. Asynchronous inference can partially mask this latency by parallelizing action execution and subsequent inference, but it introduces two critical issues: perception-execution misalignment and long reaction time. In this paper, we propose EagleVLA, a method for efficient VLA deployment on onboard devices via Foresight-Aligned Asynchronous Correction. To address misalignment, we train a lightweight future correction module that predicts future environment representations, allowing the action expert to predict actions from the future time step. To reduce reaction time, we introduce confidence-based scheduling optimization that adaptively balances VLM and action expert invocations. We also build a llama.cpp-based inference engine tailored for onboard VLA deployment, with system-level optimizations including CUDA graph reuse, GPU-resident intermediate buffering, and flow unrolling. Extensive experiments demonstrate that EagleVLA achieves 14.71x and 9.20x improvements in control frequency compared with naive PyTorch and vla.cpp on Jetson Orin, while outperforming VLASH by 14.8% in success rate on LIBERO benchmark. The code of our asynchronous algorithm is available on https://github.com/PKU-SEC-Lab/EagleVLA, and our efficient llama.cpp-based inference engine is available on https://github.com/PKU-SEC-Lab/EagleVLA-Edge.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 4 Aug 2026 · of version 4 of the paper; version 7 is newer · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A credible systems paper for edge VLA inference whose latency gains are likely real, but the success-rate story leans on an untested distribution-match assumption and a malformed confidence loss. This line is wrong

4 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the references say.

The paper says, at §4.1, Eq. (1)

The review did not quote this passage.

Pith says: Objection 1 of 4 · major

The confidence loss is malformed. L_total = L_predict + λ ||\hat c + L_predict||_2 has no ground-truth confidence target. If L_predict is a scalar, minimizing ||\hat c + L_predict||_2 simply drives \hat c toward −L_predict; it is not a calibration objective.

more of objection 1

Consequently, the confidence signal used by Algorithm 1 (lines 7–12) is not trained to predict future-prediction error, and the threshold θ (Figure 8) is unmoored. The authors need to define the target for \hat c explicitly (e.g., regress L_predict) and evaluate the calibration of the resulting confidence.

The paper says, at §4.1 and Table 3

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

Two-stage training leaves a train/test mismatch that is load-bearing. Stage 1 trains the compressor and action expert on ground-truth h_{t+Δ}; Stage 2 trains only the correction module f_corr. The action expert is never trained or fine-tuned on the predicted \hat h_{t+Δ}.

more of objection 2

Table 3 reports nearly flat success rates across Δ, but this assumes f_corr outputs lie on the GT-latent manifold. The paper provides no predicted-vs-GT latent error analysis and no oracle-latent ablation (e.g., replacing f_corr with GT h_{t+Δ} at evaluation), so the reported flat curves may not transfer to the real system. This is especially concerning in Algorithm 1: high-confidence steps update B_h with predicted latents (lines 9, 11, 21), and subsequent f_corr calls use a predicted latent as input, so errors can compound. The paper should report closed-loop drift and an oracle ablation.

The paper says, at Table 3

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

The LIBERO success rates are reported as single numbers with no error bars, seeds, or episode counts. Many differences are small (e.g., Δ=1 SPATIAL: 97.7 vs 98.8; GOAL: 97.0 vs 96.7), and the headline '14.8% average improvement' is an average over four datasets and nine Δ values. Without confidence intervals or paired comparisons across seeds, the claim is not statistically supported. The same issue affects Figure 7 (real-world trials) and Figure 8 (threshold ablation).

The paper says, at §2/Appendix B/Algorithm 1

The review did not quote this passage.

Pith says: Objection 4 of 4 · major

The conditioning mechanism is underspecified and potentially inconsistent. π0/π0.5 action experts cross-attend to layer-wise LLM KV caches (Appendix A), but the paper proposes feeding a compressed final-layer VLM state \hat h_{t+Δ} to the action expert. It is not explained how this single vector replaces or augments the layer-wise KV condition.

more of objection 4

Moreover, in Algorithm 1, the high-confidence path (lines 7–12) updates B_h but not B_kv, while the action expert consumes B_kv from an earlier VLM invocation. The combination of stale KV and predicted final-layer state is not validated. A precise architectural description and an ablation of the correction item are needed to substantiate the 'foresight-aligned' mechanism.

6 minor objections

The paper says, at Eq. (1)

Pith says: minor objection 1 of 6

If \hat c is intended to be a confidence score, its range and calibration metric should be defined; the initialization \hat c=0.0 and the condition θ<0.0 in Algorithm 1 are unexplained.

The paper says, at Table 4

Pith says: minor objection 2 of 6

The vla.cpp baseline latency of 893.0 ms is cited in the text but not included in the table; please include it for a direct comparison.

show 4 more

The paper says, at Figure 8

Pith says: minor objection 3 of 6

The x-axis includes 'w/o scheduling −0.2 −0.3 −0.4', but the choice of θ values and the variance over episodes are not reported.

The paper says, at Figure 7

Pith says: minor objection 4 of 6

The real-world results use only 10 trials per subtask and report no confidence intervals or statistical significance; a larger sample or per-trial statistics would strengthen the claim.

The paper says, at §5.1

Pith says: minor objection 5 of 6

The 'action loss' used in Stage 1 is not defined; specify the flow-matching loss and how the action expert is supervised.

The paper says, at Algorithm 1

Pith says: minor objection 6 of 6

The relationship between Δ_ae (line 9), Δ (lines 16–18), and the number of executed actions per loop iteration should be clarified; the index updates are currently ambiguous.

What would settle it

Re-run the LIBERO evaluation for Delta=5 and Delta=9 with an oracle correction item: at every point where the method would use the predicted latent \hat h_{t+Delta}, instead feed the ground-truth compressed VLM final-layer state h_{t+Delta} (computed offline by running the VLM on the future observation), keeping the scheduler, action expert, and system settings unchanged. Compare success rates and per-step latent distances (mean L2/cosine) against Table 3. If the oracle-conditioned run substantially outperforms the predicted-latent run, the train/inference distribution shift is real and the action expert must be trained on predicted latents; if they match and latent distances are small, the concern is resolved.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 4 Aug 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 1 min 19 s · 15,117 tokens read, 7,449 written
  3. Editor wrote the verdict above 21 s · 15,991 tokens read, 2,301 written
  4. Referee major revision: 4 major objections, 6 minor 50 s · 15,531 tokens read, 4,920 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Quick take: this is a solid engineering contribution with a real weak spot in the evidence for the headline success-rate curve. The latency numbers are believable; the 'foresight' correction story needs more work before I'd trust the flat accuracy claim.

What's new: the combination of a lightweight module that predicts future VLM latents conditioned on committed actions, confidence-gated skipping of the VLM, and CUDA-graph reuse/GPU-resident buffers/unrolled flow matching for VLA on Jetson. The 8.66x control-frequency improvement over naive PyTorch and 5.41x over vla.cpp are consistent with a careful roofline analysis showing both VLM and action expert are bandwidth-bound on the Orin. Real-robot results on folding clothes support the feasibility. That part deserves credit.

The soft spots are real. First, Eq. (1) is malformed: L_total = L_predict + λ || c_hat + L_predict ||_2 adds a scalar loss into the confidence norm, so the confidence head has no well-defined target. That may be a typo, but the training objective for the scheduler is literally undefined as printed. Second, the two-stage training never exposes the action expert to the correction module's predicted latents; it only trains the expert on ground-truth future states, then trains the correction module separately. At inference the expert consumes predicted latents, so unless those lie exactly on the ground-truth manifold, you have a distribution shift. The confidence-gated loop in Algorithm 1 makes this worse by feeding predicted latents back into the hidden-state buffer. The paper reports no analysis of prediction error and no oracle-latent ablation, so the flat success-rate curve in Table 3 is not backed by the experiment that would test it. Third, the LIBERO numbers have no error bars or episode counts, and some entries exceed the synchronous baseline—possible but needs comment.

None of these are fatal to the systems contribution. The latency numbers are independently plausible, and the scheduling idea is reasonable. But the headline claim that future correction solves misalignment without accuracy loss needs an ablation where the action expert is trained on predicted latents, plus a defined confidence loss and error bars. A serious referee should ask for exactly these.

I'd send this to peer review rather than desk-reject. The audience is robotics systems researchers and edge-ML people; it will be useful once the corrective evidence is added. I'd probably not cite it in its current form.

The strongest objection

The central scientific claim is that a 40M future correction module can supply the action expert with faithful future VLM latents, enabling asynchronous control without a success-rate penalty. Section 4.1's two-stage training does not establish this. Stage 1 trains the compressor and action expert on ground-truth compressed VLM final-layer states h_{t+Delta}. Stage 2 trains only the correction module f_corr to regress h_{t+Delta} from h_t and committed actions; the action expert is not subsequently trained or fine-tuned on f_corr's outputs. At inference the action expert consumes predicted latents \hat h_{t+Delta}, so unless f_corr's predictions lie exactly on the manifold of GT latents, the action expert is evaluated out-of-distribution. This is especially acute in the confidence-based scheduling loop (Algorithm 1, lines 7-12): when confidence is high, B_h is updated with the predicted latent, and the next correction step uses that predicted latent as input, so per-step errors compound. The paper reports no analysis of predicted-vs-GT latent error and no oracle-latent ablation. Additionally, the confidence loss in Eq. 1 is malformed (\|\hat c + L_predict\|_2 appears to add the regression loss inside the confidence norm), leaving the confidence head's training target undefined; without a defined target, the scheduler cannot be trusted to detect compounding drift. These are correctness risks, not consensus disagreements: the system-level latency numbers are plausible, but the evidence for the headline flat success-rate curve depends on an untested distribution-match assumption.

The reader's weakest assumption is the load-bearing one: two-stage training leaves a distribution gap between the GT future latents used to train the action expert and the predicted future latents used at inference. I agree that this gap is untested and that the reported flat success rates in Table 3 do not resolve it. The concern is sharpened by the recursive use of predicted latents in the confidence-based scheduler and by the malformed confidence loss in Eq. 1, which means the scheduler cannot be relied on to detect when predictions drift. This is an internally addressable empirical issue rather than a reason to reject the paper's system contributions: the latency and control-frequency measurements are independent of the correction module and appear credible. The reader already returned CONDITIONAL, and my analysis reinforces that verdict rather than moving it. A concrete oracle-latent comparison would settle whether the action expert's observed robustness transfers to true predicted latents or whether an additional training stage is required.

The referee's report

If the claims hold, the paper makes a useful contribution: it directly addresses an important deployment bottleneck for VLA robots, and the proposed combination of a lightweight future-latent predictor, confidence-based scheduling, and edge-oriented system optimizations is plausible and of practical interest. The release of code and the edge inference engine are also positive. However, the central scientific claim—that the correction module supplies faithful future latents so that the action expert suffers no success-rate penalty—is not fully established by the current training protocol, and the confidence mechanism is not properly defined. The latency results are promising, but the validation gaps prevent the paper from being accepted in its present form.

Pith's reading of the paper (a machine's summary, not the authors')

The paper tries to establish that vision-language-action (VLA) robot policies, which are normally too slow for low-power onboard hardware, can be run at real-time control rates by making asynchronous inference foresight-aligned. It claims that a lightweight 40M module can predict the future environment representation conditioned on the actions already committed to execution, so the action expert predicts actions from the future time step rather than from a stale observation. With that correction, plus a confidence-based scheduler and system-level optimizations, the paper reports 8.66x and 5.41x improvements in control frequency over two baseline runtimes on a low-power edge GPU, and a 14.8% average success-rate advantage over the strongest prior asynchronous method. If the claim holds, battery-powered robots could carry capable VLA policies without a major success-rate penalty.

What carries the argument. The load-bearing object is the future correction module: a roughly 40M network with a Q-Former compressor for the VLM final-layer hidden state, an MLP-plus-transformer encoder for the committed action chunk, and two output heads—a correction head that predicts the compressed future state and a confidence head that estimates its reliability. It is trained in two stages: Stage 1 teaches the action expert to consume ground-truth future states, and Stage 2 trains the module to reproduce those states from the current state and committed actions. This module carries the argument by converting the misalignment problem into a latent future-state prediction problem, and its confidence output drives t

Core claim. The central claim is that perception-execution misalignment in asynchronous VLA inference can be resolved by predicting the compressed final-layer state of the VLM at the time the next action chunk will be executed, conditioned on the actions already committed. The paper trains a 40M correction module to produce that future state from the current compressed hidden state and the committed action sequence. When the action expert conditions on this predicted future state instead of the outdated observation, task success stays nearly flat as the inference delay grows from 1 to 9 action steps, while previously proposed asynchronous methods degrade sharply. The same correction module also outputs

The premise it rests on. The load-bearing premise is that the predicted future latent is a faithful replacement for the real future environment representation; if the two diverge, the action expert is conditioning on a state it was never trained to read.

If the paper is right (Pith's inferences, not the authors')

Reaction time drops from the full policy latency to roughly the action-expert latency when confidence is high, enabling a robot to respond to environmental changes within a fraction of a second.

On the target low-power GPU, the measured control frequency rises from about 0.7 Hz to over 6 Hz, with reaction time falling from over 1.4 seconds to about 165 milliseconds.

Success rates stay near the synchronous baseline across inference delays of 1 to 9 action steps, where prior asynchronous approaches lose tens of points.

The system-level accelerations—computation graph reuse, GPU-resident intermediate buffering, and flow-matching unrolling—do not alter model computation and can be stacked with quantization or pruning.

Because the correction module is trained over a range of delay values, a single model can serve different hardware platforms or power modes without retraining.

The paper leaves implicit that the action expert is trained only on ground-truth future latents in Stage 1; fine-tuning it on predicted latents, or coupling the two modules adversarially, would close the distribution gap that could appear at real deployment.

The confidence head suggests a natural online-calibration loop: log the error between predicted and actual future representations during deployment and adjust the VLM-invocation threshold per task or per robot.

Because the correction module predicts a compressed latent rather than pixels, the approach should transfer to other VLA architectures with a similar VLM-plus-action-expert split, not just the system evaluated here.

The scheduling rationale—on bandwidth-limited edge devices the VLM and action expert are both bandwidth-bound, so parallelizing them causes contention—is a testable claim on any new edge hardware and explains why the serial, confidence-gated design wins.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper · Review history (3 versions)

The paper

Abstract

Vision-Language-Action (VLA) models have achieved impressive performance on diverse embodied tasks. However, deploying VLA models on low-power onboard devices, such as the Jetson Orin, remains challenging due to their high computational complexity, which leads to substantial inference latency and low control frequency. Asynchronous inference can partially mask this latency by parallelizing action execution and subsequent inference, but it introduces two critical issues: perception-execution misalignment and long reaction time. In this paper, we propose EagleVLA, a method for efficient VLA deployment on onboard devices via Foresight-Aligned Asynchronous Correction. To address misalignment, we train a lightweight future correction module that predicts future environment representations, allowing the action expert to predict actions from the future time step. To reduce reaction time, we introduce confidence-based scheduling optimization that adaptively balances VLM and action expert invocations. We also build a llama.cpp-based inference engine tailored for onboard VLA deployment, with system-level optimizations including CUDA graph reuse, GPU-resident intermediate buffering, and flow unrolling. Extensive experiments demonstrate that EagleVLA achieves 14.71x and 9.20x improvements in control frequency compared with naive PyTorch and vla.cpp on Jetson Orin, while outperforming VLASH by 14.8% in success rate on LIBERO benchmark. The code of our asynchronous algorithm is available on https://github.com/PKU-SEC-Lab/EagleVLA, and our efficient llama.cpp-based inference engine is available on https://github.com/PKU-SEC-Lab/EagleVLA-Edge.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-23

The authors' main claim, as the classifier picked it out 0.99

“Extensive experiments demonstrate that Jetson-PI achieves 8.66x and 5.41x improvements in control frequency compared with naive PyTorch and vla.cpp on NVIDIA Jetson Orin, while outperforming VLASH by 14.8\% in average success rate on the LIBERO benchmark.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.14 · human participants 0.03 · controlled comparison 0.95 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.99

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v5). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Vision-Language-Action (VLA) models have achieved impressive performance on diverse embodied tasks. background 1.00 0.05 0.02 asserted 1.00
However, deploying VLA models on low-power onboard devices, such as the Jetson Orin, remains challenging due to their high computational complexity, which leads to substantial inference latency and low control frequency. background 0.85 0.10 0.02 asserted 1.00
Asynchronous inference can partially mask this latency by parallelizing action execution and subsequent inference, but it introduces two critical issues: perception-execution misalignment and long reaction time. background 0.80 0.11 0.03 asserted 0.82
In this paper, we propose Jetson-PI, a method for efficient VLA deployment on onboard devices via Foresight-Aligned Asynchronous Correction. contribution 0.99 0.85 0.15 asserted 0.98
To address misalignment, we train a lightweight future correction module that predicts future environment representation conditioned on committed actions, enabling the action expert to directly predict actions from the future time step. method 0.92 0.92 0.04 asserted 1.00
To reduce reaction time, we introduce confidence-based scheduling optimization that adaptively balances VLM and action expert invocations, complemented by system-level accelerations including CUDA graph reuse, GPU-resident intermediate buffering, and flow unrolling. method 0.92 0.93 0.05 asserted 1.00
Extensive experiments demonstrate that Jetson-PI achieves 8.66x and 5.41x improvements in control frequency compared with naive PyTorch and vla.cpp on NVIDIA Jetson Orin, while outperforming VLASH by 14.8\% in average success rate on the LIBERO benchmark. result 1.00 0.97 0.76 asserted 1.00
The code of our asynchronous algorithm is available on https://github.com/PKU-SEC-Lab/Jetson-PI, and our efficient llama.cpp-based inference engine is available on https://github.com/PKU-SEC-Lab/Jetson-PI-Edge. contribution 0.76 0.64 0.02 asserted 0.99

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 1.00

What was examined, and what was not

Sources read

  • arXiv listing abstract v5

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,584 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Cited by 5 Pith papers
How the literature uses it

What this paper does with the arXiv papers it cites

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

What it cites · 42 sentences without a confident label

Uses data or values from 1 paper
Compares with 2 papers
Cites as support 2 papers
Cites as background 22 papers
  • As shown in Figure [ref], asynchronous inference eliminates pauses between chunks by parallelizing execution of the current chunk with inference of the next chunk [ref] .

    2509.23224 · Background · 0.82

  • As shown in Figure [ref], asynchronous inference eliminates pauses between chunks by parallelizing execution of the current chunk with inference of the next chunk [ref] .

    World Action Models are Zero-shot Policies · Background · 0.82

  • As shown in Figure [ref], asynchronous inference eliminates pauses between chunks by parallelizing execution of the current chunk with inference of the next chunk [ref] .

    2512.24673 · Background · 0.82

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment