Skip to main content

Deploying VLAs on G1: Architecture & Integration

Duration: 65 min · Level: Advanced · Module: 5. Foundation Models & VLA Architecture · Focus: deployment, inference, hardware, architecture

Learning objectives

By the end of this lesson you will be able to explain and apply:

  • NVIDIA AGX Orin
  • Two-level control architecture
  • Quantization
  • Speculative decoding
  • Multi-GPU setup

Why this matters​

Running a 3-7B parameter VLA model at useful frequency on a battery-powered humanoid requires careful co-design of the inference pipeline, compute allocation, and control architecture.

Overview​

Running a 3-7B parameter VLA model at useful frequency on a battery-powered humanoid requires careful co-design of the inference pipeline, compute allocation, and control architecture. This lesson covers the practical engineering of deploying foundation models on G1.

Key concepts​

Key idea

NVIDIA AGX Orin: 275 TOPS INT8, 64GB LPDDR5; can run 7B parameter quantized model at ~5 Hz; sufficient for high-level task planning but not high-frequency control

  • Two-level control architecture: VLA at 5-10 Hz for language-conditioned task planning → low-level reactive controller at 1-2 kHz for joint execution
  • Quantization: INT4/INT8 quantization reduces model size 2-4×, inference speed 2-3×; minimal quality loss for VLA action prediction with AWQ quantization
  • Speculative decoding: generate action candidates in parallel; evaluate with discriminator; reduces effective latency by 2-3× for transformer-based policies
  • Multi-GPU setup: Figure 02 uses 2× Orin NX modules — one for perception + VLA inference, one for locomotion control and safety monitoring
  • Safety wrapper: VLA outputs pass through a safety filter that checks joint limits, velocity limits, and collision prediction before execution — essential for healthcare

Check your understanding​

Try to recall each answer before expanding it.

Q1. What do you know about NVIDIA AGX Orin?

275 TOPS INT8, 64GB LPDDR5; can run 7B parameter quantized model at ~5 Hz; sufficient for high-level task planning but not high-frequency control

Q2. What do you know about Two-level control architecture?

VLA at 5-10 Hz for language-conditioned task planning → low-level reactive controller at 1-2 kHz for joint execution

Q3. What do you know about Quantization?

INT4/INT8 quantization reduces model size 2-4×, inference speed 2-3×; minimal quality loss for VLA action prediction with AWQ quantization

Q4. What do you know about Speculative decoding?

generate action candidates in parallel; evaluate with discriminator; reduces effective latency by 2-3× for transformer-based policies

Q5. What do you know about Multi-GPU setup?

Figure 02 uses 2× Orin NX modules — one for perception + VLA inference, one for locomotion control and safety monitoring

← Previous: 5.4 Diffusion Policy — Visuomotor Control via Denoising

Part of Module 5: Foundation Models & VLA Architecture.