## Does RL Fine-Tuning Solve the VLA Latency Problem Killing Real-World Deployment?

Yes — and a new Stanford paper puts numbers to it. Researchers Perry Dong, Kuo-Han Hung, Dorsa Sadigh, and Chelsea Finn report that their Real-Time EXPO-FT framework lifts average policy performance from **42% to 97%** across four dynamic real-world manipulation tasks, using no more than **10 minutes of online robot data** and zero human intervention during the adaptation process. The core insight is architectural: decouple a large, slow [Vision-Language-Action model](https://humanoidintel.ai/glossary/vision-language-action-model) that proposes action chunks from a lightweight, fast "edit policy" that reactively corrects those actions using the latest observation. Reinforcement learning then fine-tunes the full system in a way that [imitation learning](https://humanoidintel.ai/glossary/imitation-learning) alone cannot — pushing the policy beyond its training distribution toward genuinely reliable performance. On the Kinetix simulation benchmark, the method achieves the best performance among both delayed and non-delayed methods in **10 out of 10 environments**. For anyone building humanoid manipulation pipelines on top of foundation VLA models, this is one of the more concrete results published this year.

---

## The Core Problem: Inference Latency Corrupts VLA Reliability

The dirty secret of deploying large VLA models on physical hardware is timing. Because modern VLA models are computationally heavy, the camera observation used to compute an action is often substantially older than the robot's current state by the time that action executes. The paper's authors frame this as a **distribution shift**: the policy was trained on fresh observations, but at inference time it acts on stale ones. In dynamic manipulation tasks — catching a passed object, balancing a ball, kicking in table soccer, or picking a moving target — this staleness compounds rapidly into failure.

Prior approaches attacked this through asynchronous policy execution, running inference and motor control on separate threads to minimize blocking. But as the authors note, those methods are almost exclusively built on imitation learning. They can reduce the *effect* of latency but have no principled mechanism for improving *beyond* the training distribution. When the task dynamics exceed what the demonstration data covered, the policy degrades and there is no recovery path.

This is not a theoretical concern. As humanoid platforms like [Figure AI](https://humanoidintel.ai/companies/figure-ai)'s deployments and others push toward dynamic, contact-rich tasks in unstructured environments, the gap between benchmark VLA performance and real-world reliability has remained a persistent, publicly acknowledged problem across the industry.

---

## The Real-Time EXPO-FT Architecture

The paper builds on **EXPO-FT**, an existing framework for sample-efficient VLA fine-tuning with reinforcement learning. Real-Time EXPO-FT extends it with a two-tier architecture:

**1. The slow tier — action chunk proposal:** A large, pretrained VLA generates action chunks. This is the computationally expensive step, and it runs asynchronously. The VLA brings strong behavioral priors from pretraining — generalizable task understanding that would be prohibitively expensive to learn from scratch.

**2. The fast tier — the edit policy:** A lightweight module runs at the cadence required for real-time control. Conditioned on the *latest* observation (not the stale one the VLA saw), the edit policy modifies the proposed action chunk before execution. It is reactive rather than generative — small, fast corrections rather than full action synthesis.

**3. RL as the training signal:** Critically, the entire system is fine-tuned with reinforcement learning rather than behavioral cloning. This allows the policy to discover action corrections that exceed what any human demonstration could have covered, and it does so without requiring human intervention during the online adaptation phase.

The four real-world tasks chosen — robot object passing, ball balancing, table soccer kicking, and dynamic object picking — are specifically selected to stress-test the latency problem. All involve fast-changing states where a stale observation is particularly costly. The reported improvement from 42% to 97% average performance using a 10-minute online data cap is the headline result, but the benchmark result (best in class across 10/10 Kinetix environments, against both delayed and non-delayed baselines) is arguably the more rigorous validation.

---

## Why This Matters for Humanoid Stack Architects

The implications here extend well past academic benchmarking. The industry's current trajectory is to fine-tune large foundation VLA models — the kind being developed or deployed by labs affiliated with [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence), [Skild AI](https://humanoidintel.ai/companies/skild-ai), and others — on specific robot hardware and task distributions. The standard approach has been [imitation learning](https://humanoidintel.ai/glossary/imitation-learning) from human teleoperation data. That approach is expensive, slow, and bounded by the quality and coverage of the demonstration set.

Real-Time EXPO-FT suggests a different path: use a strong pretrained VLA as a prior, then let RL handle the hardware-specific, latency-sensitive adaptation — automatically, in minutes rather than days of data collection. The 10-minute cap on online robot data is not a limitation dressed up as a feature; it's the point. If the system can reach 97% average task performance autonomously in that window, the economics of deploying and re-deploying humanoid manipulation policies change substantially.

**The skeptical read:** The tasks demonstrated, while dynamic, are constrained laboratory setups. Ball balancing and table soccer kicking are compelling demonstrations but do not capture the full complexity of [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) in unstructured environments — variable object geometries, cluttered scenes, or the multi-step tool-use tasks that commercial humanoid deployments actually demand. The method also inherits whatever capability ceiling the underlying pretrained VLA has. If the VLA cannot represent an action class, the edit policy cannot invent it. The 42%-to-97% improvement is real, but it measures improvement *within* the task distribution of dynamic-but-bounded lab tasks.

That said, the sample efficiency result is difficult to dismiss. Ten minutes of autonomous robot interaction to achieve near-ceiling performance on challenging dynamic tasks is a qualitatively different operational profile than the days-long teleoperation campaigns that today's fine-tuning pipelines require.

---

## Industry Trajectory

The broader pattern here is the convergence of three trends: (1) large pretrained VLAs becoming the default starting point for manipulation policy development; (2) RL regaining credibility as a fine-tuning tool after years of being overshadowed by [imitation learning](https://humanoidintel.ai/glossary/imitation-learning) in robotics; and (3) latency as the next major systems engineering bottleneck as these models scale. Real-Time EXPO-FT addresses all three simultaneously.

For engineering teams at humanoid companies evaluating their AI stack, the immediate question is whether the EXPO-FT framework generalizes to their specific VLA backbone and hardware. The paper's authors are from Stanford's Sadigh and Finn labs — groups with strong track records of work that translates, but also work that requires non-trivial engineering effort to replicate outside of the originating lab's setup.

The Kinetix benchmark performance and the real-world task results together make this one of the stronger RL-for-VLA papers of the year. The code and methodology details at the paper's website will determine how quickly the field can build on it.

---

## Key Takeaways

- **42% → 97%:** Real-Time EXPO-FT improves average policy performance across four dynamic real-world tasks using a maximum of 10 minutes of online robot data and no human intervention.
- **10/10 on Kinetix:** The method outperforms both delayed and non-delayed baseline methods across all 10 Kinetix benchmark environments.
- **Two-tier architecture:** A slow, expressive VLA generates action chunks; a fast, lightweight edit policy reactively corrects them using current observations — decoupling the latency problem from the expressiveness requirement.
- **RL over imitation learning:** The RL fine-tuning signal is what allows the policy to exceed its training distribution, something imitation-learning-only approaches cannot do.
- **Stack implications:** Sample-efficient autonomous fine-tuning could substantially change the economics of deploying and redeploying humanoid manipulation policies in the field.
- **Skeptic's note:** Demonstrated tasks are dynamic but lab-bounded; generalization to full commercial manipulation complexity remains to be validated.

---

## Frequently Asked Questions

**What is the VLA latency problem in humanoid robotics?**
Large Vision-Language-Action models take significant time to compute an action from a camera observation. By the time the action executes, the robot's environment has changed, creating a mismatch between what the policy "saw" and the current state. In dynamic tasks, this staleness directly causes failures.

**What is Real-Time EXPO-FT and how does it work?**
Real-Time EXPO-FT is a reinforcement learning framework from Stanford researchers (Perry Dong, Kuo-Han Hung, Dorsa Sadigh, Chelsea Finn) that fine-tunes VLA policies for real-time control. It uses a large VLA to propose action chunks asynchronously, while a lightweight "edit policy" reactively adjusts those actions based on the latest observation. RL trains the combined system to exceed imitation learning performance bounds.

**How much data does Real-Time EXPO-FT require?**
The paper reports achieving its results with online robot data capped at 10 minutes per task, with no human intervention required during the adaptation process.

**How does this compare to existing asynchronous VLA methods?**
Prior asynchronous methods reduce latency effects but are built on imitation learning, which limits them to the training distribution. Real-Time EXPO-FT uses RL, enabling the policy to improve beyond what demonstrations covered.

**What tasks did the researchers use to validate the approach?**
Four real-world dynamic manipulation tasks: robot object passing, ball balancing, table soccer kicking, and dynamic object picking. These were chosen specifically because fast-changing states make observation staleness especially costly.