Seonglae ChoVision Language Action Model
Should have action decoder or low-level vector-based controlling. If VLA generates just a high-level sequence, it is VLM not VLA
Vision Language Action Models
Vision Language Action Notion
Steering
VLA's Transformer FFN neurons still maintain semantic concepts like slow, fast, up. By selectively activating these neurons (activation steering) → robot behavior can be adjusted in real-time without fine-tuning, rewards, or environment interaction. In both simulation (OPENVLA, LIBERO) and real robots (UR5, Pi 0) → behavioral characteristics like speed and movement height change in a zero-shot manner. Semantic-based neuron intervention is more effective than prompt modification or random intervention. VLAs maintain interpretable semantic structures internally, which can be directly manipulated to control robot behavior transparently and immediately.
arxiv.org
https://arxiv.org/pdf/2509.00328
SteerVLA
Non-real-time steering through language-based policy conditioning. In other words, conditioned policy training via learned prompt interface
arxiv.org
https://arxiv.org/pdf/2602.08440
Physical Intelligence
Extract the key information needed for fine-grained manipulation from the internal features of a frozen VLA model and convert it into a compact readout representation called an “RL token” (). Using this token as input, add a lightweight actor–critic head that selects actions, and train it with online RL:
Here, is the learned RL policy, and serves as a regularizer that encourages the policy to follow the frozen VLA’s behavior, improving exploration efficiency and preventing catastrophic forgetting / knowledge collapse.
Combine the extracted RL token with the robot’s current proprioceptive state to output a 10-step action chunk (action chunking). Action chunking mitigates the temporal credit-assignment problem, accelerating learning even under sparse rewards, and enables stable control at high control frequencies (e.g., 50 Hz).
arxiv.org
https://arxiv.org/pdf/2604.23073