Why Adding an Action Expert Can Hurt, and How to Prevent It | Adople AI Research

Adople builds enterprise AI solutions and AI agents that automate critical workflows, connect fragmented data, and transform information into intelligent action.

  • Address

    600 California St, San Francisco, CA 94108
  • Email

    info@adople.ai
  • Contact

    +1 (415) 630-2010
Robotics AIResearch Note · VLAsNeurIPS 2025

Why Adding an Action Expert Can Hurt, and How to Prevent It

In this article, you will learn why adding a robot “motor control” module to a vision-language model can damage what the model already knows, how knowledge insulation prevents that, and what the authors measured when they compared it with earlier designs.

Topics we will cover
  • How vision-language-action (VLA) models evolved through two generations
  • Why gradients from an action expert harm the pretrained backbone
  • The three ingredients of knowledge insulation: stop-gradient, FAST action tokens, and co-training on web data
  • The results on training speed, inference speed, language following, and generalization
  • What engineers can take from this when building multimodal systems
Recency note: This article describes research published in May 2025 and presented at NeurIPS 2025, as of November 2026. It does not cover later models from the same group.
01Introduction

Why Robots Borrow From Language Models

Vision-language models (VLMs) are trained on web-scale data, so they carry broad world knowledge and can reason about what they see. The idea behind a vision-language-action (VLA) model is to start from such a VLM and teach it to control a robot. The robot then inherits semantic understanding, such as recognizing objects it never saw in robot training data.

The catch is that a VLM only produces discrete tokens (text), while a robot needs precise, continuous motor commands. The authors use a brain analogy. The VLM’s vision encoder is like a visual cortex, its language model is like a prefrontal cortex, and a VLA needs a third piece, a “motor cortex.” In AI the order is backwards compared with biology. We start with language, add vision, and only then bolt on motor control. The question is how to attach it without damaging the capabilities already there.

Visual cortex
Vision encoder · sees
Prefrontal cortex
Language model · reasons
Motor cortex
Action expert · moves
02Background

Two Generations of VLAs

First generation (RT-2, OpenVLA). These models output actions as tokens. Each robot joint angle is split into fixed-size bins, and each bin gets a token, so predicting an action looks like answering a question with numbers. This works for basic tasks such as picking up objects. The authors report that it is poorly suited to precise, fluent, high-frequency motion, and that it is slow at runtime. Their analogy is controlling your arm by verbally naming which muscles to contract. The paper also notes that autoregressive VLAs of this type typically run at control frequencies below 2 Hz.

Second generation (π0 and similar). These models add a new neural module that produces continuous actions, usually with diffusion or flow matching. The module is called an action expert or an action head. It can read the VLM backbone’s internal representations but specializes in motor control. This enables much more complex tasks, such as folding laundry or making a bed.

FIG. 01 — EVOLUTION OF VLA ARCHITECTURES
Three ways to get actions out of a vision-language model
GEN 1 · RT-2 / OpenVLA

Actions as tokens

cross-entropy∇ intobackbonediscrete action tokens···binned joint valuesVLM backbone(3B)Inputs: camera images · instruction · robot state
Example:RT-2, OpenVLA
Trade-off:Keeps VLM knowledge. Slow at runtime (typically under 2 Hz for autoregressive actions).
GEN 2 · π0; action-head designs

Continuous actions with an action expert

flow matching /diffusioncontinuous actionsVLM backbone(3B)representations disturbedActionexpertexpert readsbackbone∇ from new, untrained modulereaches backboneInputs: camera images · instruction · robot state
Example:π0; action-head designs
Trade-off:Fast, continuous actions. Slower training and knowledge loss reported.
GEN 3 · π0.5 + KI

Protect backbone knowledge using stop-gradient

cross-entropy∇ backbone'sown signaltoken headdiscrete FAST tokensVLM backbone(3B)expert reads backbonestop-gradientflow matching /diffusioncontinuous actionsActionexpertInputs: camera images · instruction · robot state
At inference the tokens are discarded; only the expert runs.
Example:π0.5 + KI
Trade-off:Fast continuous actions. Trains quickly. Pretrained knowledge is protected.
03The Engineering Problem

Knowledge Insulation

When you attach an action expert to a pretrained VLM and train both end to end, the backbone receives two kinds of gradients. One comes from whatever language or token loss you keep. The other comes from the expert’s continuous-action loss, and that one is the problem. The expert’s weights are newly added and untrained, and the authors report that their gradients create unfavorable learning dynamics in a backbone whose representations were tuned on web-scale data.

The authors report two primary effects:

Slow
Training becomes significantly slower (up to 7.5x slower steps).
Eroded
The backbone loses pretrained knowledge, and language following degrades.

Their illustration is a spoon-into-dish-container task, where the naive model grasps the trash instead. They offer a hypothesis, not a proven mechanism. The pretrained backbone attends to language well, the expert’s gradients interfere with that, and the model picks up other correlations in the robot data first.

FIG. 02 — KNOWLEDGE-INSULATED ACTION EXPERT
GRADIENT CONTROL
cross-entropy losstoken headVLMbackbone(3B)representations disturbedexpert reads backbonerepresentations∇ flow-matching lossreaches backboneflow-matching losscontinuous actionsActionexpert(300M)noiseInputs: camera images · prompt · robot state
cross-entropy losstoken head∇ cross-entropy(FAST tokens + text)VLMbackbone(3B)expert reads backbonerepresentationsstop-gradientflow-matching losscontinuous actionsActionexpert(300M)noiseInputs: camera images · prompt · robot state
Legend:
Forward pass (computation flow)
Backward pass (gradient flow)
Expert’s gradient (from flow-matching loss)
Backbone’s own gradient (from cross-entropy loss)
Where the expert’s gradient is allowed to go.
04The Design

Three Mechanisms

Knowledge insulation (KI) combines three changes. The first two form a chain: stopping the gradient protects the backbone but cuts off its robot signal, and a token loss supplies that signal. The third adds more training data to reinforce the backbone’s knowledge.

4.1 Stop the gradient.

The expert still reads the backbone’s representations, but its loss cannot update the backbone. This fully insulates the backbone. By itself it is not enough. The backbone now never sees robot data, so its representations are not shaped for motor control, and the expert has a harder time producing correct motions. The authors show this in their comparisons.

4.2 Give the backbone its own robot-relevant objective.

The backbone is trained with next-token prediction on discretized actions in the FAST format. The expert’s loss no longer carries robot signal into the backbone, and this loss takes over that job. The authors report that FAST tokens make the model acquire motor representations significantly faster, and that a cross-entropy loss interferes much less with pretrained knowledge than the expert’s gradients do.

FAST is a compression-based tokenization scheme for robot actions, built on the discrete cosine transform (see the FAST paper in Sources). Its tokens give a better learning signal than simply binning each joint value.

4.3 Co-train on other next-token data.

The backbone already has a token loss, so more token tasks can be added without changing the architecture. The model is trained on the full π0.5 mixture, which includes general vision-language data from the web and high-level robot commands (planning data). The goal is to keep reinforcing the semantic knowledge in the backbone.

Table 2: Loss Routing Across Data Streams
Data StreamLoss TypeParameters Receiving Gradient
Robot trajectories, as FAST tokensCross-entropy (next token)Backbone
Web vision-language dataCross-entropy (next token)Backbone
High-level robot commands (planning)Cross-entropy (next token)Backbone
Robot trajectories, as continuous actionsFlow matchingAction expert only

4.4 One model, two paths.

FIG. 03 — TRAINING & INFERENCE FLOW
π0.5 + KI
TRAINING
Inputs:camera images ·prompt ·robot stateVLMBackbone(3B)token head(FAST + text)cross-entropy loss(FAST tokens + text)expert reads backbonerepresentationsActionexpert(300M)noiseflow-matching loss(continuous actions)
INFERENCE
Inputs:camera images ·prompt ·robot stateVLMBackbone(3B)token head(FAST + text)unused: tokens discardedexpert reads backbonerepresentationsAction expert(300M)noiseiterativeflow-matchingstepscontinuousactionchunk
The token path exists only to teach the backbone. Only the expert runs at deployment.
05Results

Results: Training Speed, Execution Time & Generalization

Training speed.

On the bussing task with a generalist model, π0.5 + KI needed 7.5 times fewer training steps than π0 to reach similar performance, and it trained about as fast as π0-FAST.

Execution time.

On the single-embodiment bussing task, π0-FAST needed about twice the time of π0.5 + KI to finish, because of autoregressive inference. π0.5 + KI had the highest performance and followed language well.

Fig. 04 — Illustrative task-completion time · bussing (single embodiment)
π0.5 + KI1.0× · fastest
1.0×
π0 (baseline)~1.1×
~1.1×
π0-FAST (autoregressive)~2.0×
~2.0×
⏱ Why the gap?

π0-FAST decodes actions token-by-token, so bussing takes about twice as long. π0.5 + KI discards FAST tokens at inference and flows continuous actions through the small expert — highest performance with strong language following.

7.5×
fewer steps vs π0
2×
faster than FAST
What each comparison isolates
VariantHow it differs from π0.5 + KI
π0Naive action expert; gradients flow into the backbone
π0-FASTNo action expert; actions are autoregressive tokens
Joint trainingToken and expert losses together, with no stop-gradient
Frozen backboneThe backbone gets no robot training
π0.5 + KI without VLM dataSame recipe, without co-training on VLM (web) data

Manipulation tasks & Object Generalization

Shirt folding: π0.5 + KI had the highest average success. π0-FAST was close behind, and its printed p-value (p=0.782) shows no meaningful difference from π0.5 + KI. π0 (p=0.002) and the frozen backbone (p=0.001) were clearly lower. Freezing the backbone is not workable for this task.

Items in drawer (static robot, unseen environment): π0.5 + KI had the highest task completion. The printed p-values were 0.049 for joint training and 0.030 for π0-FAST. For π0 and the frozen backbone they were below 0.001.

Fig. 05 — Object generalization · mobile manipulator · unseen objects & envs
Without VLM data
in-dist + OOD language & performance
π0.5 + KI82%
Joint training38%
KI >> joint on all 4 metrics
With VLM data
in-dist + OOD language & performance
π0.5 + KI86%
Joint training83%
similar · overlapping error bars
ℹ️
Reading: Co-training on web data contributed the most. Stop-gradient mattered most when that data was absent.
06Practice

Practical Takeaways for Engineers

These are interpretations, not claims from the paper.

01
Detach what a new head reads.
If you attach an untrained module to a pretrained model, cutting its gradient into the pretrained weights is a cheap first protection.
02
Replace the signal you cut.
A frozen or fully detached backbone cannot adapt to the new domain, and the frozen-backbone variant performed poorly on the manipulation tasks. Give the backbone its own objective for the new modality.
03
Train with one representation, serve with another.
The discrete tokens help only during training, and the continuous head runs at inference.
04
Keep general data in the mix.
On the object-generalization task, co-training on web data had the largest effect.
05
Check significance before claiming a win.
On shirt folding, the leading baseline is statistically indistinguishable from π0.5 + KI.
07Limits

Limitations

  • —The results come from specific robot tasks (bussing, shirt folding, items in drawer, object grasping). Do not generalize them to all manipulation.
  • —The explanation for the interference is stated as a hypothesis.
  • —This article covers the May 2025 work as of November 2026. It does not cover later models.
08Outlook

Where This Is Going

The authors expect the next generation of VLAs to integrate web-scale knowledge with motor control more deeply, with more sophisticated sequential reasoning, planning, and deliberate, goal-directed task execution. They describe VLAs as among the most multimodal models in existence, because of the variety of robot sensors and actuators, and they treat training them as a branch of multimodal machine learning. They point to a better understanding of learning dynamics as a key step toward that.

Conclusion
In this article, you learned why attaching a continuous-action module to a pretrained vision-language model can slow training and erode knowledge, and how knowledge insulation addresses this. The recipe has three parts: block the action expert’s gradients, teach the backbone with discrete FAST tokens, and keep co-training on web and planning data. The reported result is a model that needed 7.5 times fewer training steps than π0, ran about twice as fast as π0-FAST on the bussing task, and kept stronger language following and generalization.
Sources
🔗
Physical Intelligence research post: “VLAs that Train Fast, Run Fast, and Generalize Better” (May 28, 2025): https://www.pi.website/research/knowledge_insulation
🔗
Driess et al. arXiv Preprint & NeurIPS 2025 Paper: “Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better,” arXiv:2505.23705 (May 29, 2025): https://arxiv.org/abs/2505.23705
🔗
Open-source model weights (openpi): https://github.com/Physical-Intelligence/openpi
🔗
FAST Action Tokenization: “FAST: Efficient Action Tokenization for Vision-Language-Action Models”: https://arxiv.org/abs/2501.09747