Embodied AI and Robot Learning: Vision-Language-Action Models — scaling robot manipulation via learning from diverse demonstrations

Embodied AI—autonomous agents perceiving and acting in physical environments—requires learning sensorimotor policies (visual input → action output) from demonstrations. RT-2 (Robotics Transformer 2, Google DeepMind, 2023) demonstrates that vision-language models fine-tuned on robot trajectories generalize across tasks and embodiments.

Visuomotor Policy Architecture

Policies learn direct visual-to-action mapping: images (RGB camera) → end-effector pose, gripper state. Convolutional encoder (ResNet) extracts visual features; recurrent modules (LSTM, temporal attention) maintain action history; action decoder outputs normalized motor commands (position, velocity, gripper). Training: behavioral cloning (imitation learning) from human demonstrations via supervised learning.

RT-2 and Vision-Language Foundation Models

RT-2 leverages pre-trained vision-language models (VLM: image + text → text generation). Fine-tuning tokens: vision encoder (frozen or trainable), language model (frozen), task-specific adapter. Clever insight: reframe robot action as text generation. Image→VLM tokenizes visual observations, language model predicts tokens corresponding to actions (e.g., move forward 10cm → token representation). Transfer: model learned to predict actions generalizes to novel objects, scenes, and tasks.

Behavior Cloning and Demonstration Collection

RT-2 trained on 11M robot trajectories from 13 robots across diverse tasks (pick, place, push, wipe). Behavioral cloning: minimum supervised loss between predicted and ground-truth actions. No reward signal required—direct imitation. Challenges: distribution shift (model's errors compound in open-loop execution), multi-modal actions (multiple correct responses to same image).

Sim-to-Real Transfer and Domain Randomization

Simulation (MuJoCo, Gazebo, CoppeliaSim) enables cheap data collection (no robot hardware wear, faster iteration). Domain randomization (random textures, lighting, object sizes, physics parameters) trains simulation policies to be robust to visual/dynamics variation. Transfer to real robots often succeeds with minimal fine-tuning. Physics engine fidelity (contact dynamics, friction) impacts transfer quality.

DROID and ALOHA Datasets

DROID (Distributed Robotics Open Interactive Dataset): 2.1M trajectories from 11 universal robots, open-source. ALOHA (A Low-cost Open-source maniPulator with High-resolution vIsion): teleoperated bimanual arm with synchronized manipulation recorded in real homes/offices. These large-scale datasets enable scaling robot learning, moving toward foundation models for robotics.