Embodied VLAs — RT-2 to GR00T, the action-format arc
four action formats in use today
discrete 256-bin
RT-2 (2023), OpenVLA
one token per DOF
text-vocab compatible
pros: simple
cons: ~300 tok/s needed
inference 3-5 Hz
FAST (DCT + quantize)
2024 improvement
~10 tokens / trajectory
low-freq coefficients
pros: 3-5x faster
cons: loses high-freq
used by pi0-FAST
flow-matching head
pi0 (2024)
continuous output
50-step trajectory
pros: smoothness
5 denoise steps
~50 Hz control
dual-system
GR00T N1 (2025)
System 2: VLM ~1 Hz
System 1: small ~100 Hz
pros: humanoid-scale
subgoals + fast control
best for 30+ DOF
training recipe + safety
co-fine-tune
ratio ~0.5:1 to 1:1
web VQA + robot demos
preserves general knowledge
robot-only -> forgets lang
fine-tune path
LoRA on 100-1000 demos
to adapt to new robot
Open X-Embodiment
22 datasets
1M trajectories
22 robot embodiments
ALOHA / Bridge / Droid
unified schema
state, camera, action
open-source
safety gates (outside VLA)
hard joint limits
velocity clipping
workspace bounds
HITL approval for novel tasks
VLA suggests, controller enforces
always treat VLA output as prior,
not a command