Open-Source VLA Models: Four Factions Battling for the Robot Brain and a Trillion-Dollar Market

Forum topic · 小凯 · 2026-03-29

Summary

This in-depth analysis examines the wave of open-source Vision-Language-Action (VLA) models for robotics, mapping the ecosystem into four factions: academic projects (OpenVLA, Octo), tech giants building ecosystems (NVIDIA GR00T N1, Google Gemini Robotics), rising Chinese players (Xiaomi-Robotics-0, Ant Group's LingBot-VLA, Unitree UnifoLM-VLA-0), and performance-focused teams like Physical Intelligence's π0/π0.5 series. Key technical comparisons cover dual visual encoders, diffusion policies, flow matching, discrete action tokens, and the control-frequency arms race from 10Hz (RT-2, OpenVLA) to 200Hz+ (Figure's Helix). The article dissects the business logic behind 'free' robot brains—ecosystem lock-in (NVIDIA), data flywheels (Google), standards-setting, and hardware-driven traffic—contrasting genuine open source with ecosystem-locking pseudo-open strategies. It concludes that open source can compete on models via a three-layer flywheel (models + data + tools), but remains behind giants on proprietary data and hardware, with the most likely outcome (50% probability) being coexistence: open source dominating research, closed systems dominating commercial deployment.

Vision-Language-Action Models: Technical Deconstruction and Business Anatomy

In early 2026, a quiet revolution is unfolding in robotics. Google, NVIDIA, Unitree, Ant Group, and Xiaomi—apparently unrelated tech giants—nearly simultaneously opened up their robot "brains" for free. This is not charity but a strategic battle over a trillion-dollar market. This article breaks down the four major technical factions, the business logic behind open-sourcing, and how the open-vs-closed route debate will shape the future of embodied intelligence.

Key points

Four factions now define the open-source VLA ecosystem: academic idealists (OpenVLA, Octo), giant ecosystem players (NVIDIA GR00T, Google Gemini Robotics), Chinese challengers (Xiaomi, Ant/LingBot, Alibaba RynnVLA, Unitree UnifoLM-VLA, Qianxun Spirit v1.5), and the performance-focused π0 series from Physical Intelligence.
OpenVLA (7B) beat Google DeepMind's RT-2-X (55B) by 16.5% across 29 manipulation tasks (June 2024), thanks to a dual visual encoder design: DINOv2 for spatial/geometric understanding, SigLIP for semantics, fused by Llama 2 7B, trained on Open X-Embodiment (71 datasets, 1M+ trajectories), fully open under Apache 2.0.
Octo is the lightweight "Swiss army knife": diffusion transformer, modular inputs/outputs, only 27M–93M parameters, trained across 25 robot platforms and 800k trajectories.
NVIDIA GR00T N1→N1.6 is not just a model but a full ecosystem: Thor/Jetson chips, Isaac Sim/Lab simulation, OSMO data platform, free model, paid NIM inference—classic ecosystem lock-in.
Google Gemini Robotics bets on end-to-end multimodal capability, real-world data feedback into Gemini, and owning the robot-era interaction entry point.
Chinese entrants: Xiaomi-Robotics-0 (Feb 2026) targets consumer home robots with an asynchronous execution architecture and action-chunk time alignment; Ant's LingBot-VLA (Jan 2026) focuses on service robots with a full toolchain; Unitree's UnifoLM-VLA-0 supports multiple robots (G1, H1, Go2) with a single policy covering 12 task categories.
Physical Intelligence π0/π0.5: flow matching + MoE, 50Hz control; π0 achieved laundry folding; π0.5 introduced "Knowledge Insulation" (joint training on robot + web + language data) and chain-of-thought reasoning, achieving >90% success on unseen households with zero-shot generalization.

Architecture and Technical Routes

| Model | Architecture | Params | Core innovation | |---|---|---|---| | OpenVLA | Prismatic VLM (dual encoders) | 7B | Dual vision encoders | | Octo | Diffusion Transformer | 27M–93M | Diffusion policy, modularity | | π0/π0.5 | Flow Matching + MoE | 5B+ | Continuous actions, knowledge insulation | | GR00T N1 | Transformer | undisclosed | Deep NVIDIA toolchain integration | | Helix (Figure AI) | Dual-system | undisclosed | System 1/System 2 split |

Action representation philosophies:

Discrete tokens (OpenVLA, RT-2): discretize actions into 256 bins; stable training, reuses LLM weights, but quantization limits fine control.
Diffusion (Octo, RDT): models multi-modal action distributions and smooth motion, but slow multi-step inference.
Flow matching (π0): elegant continuous mapping from noise to action, high sampling efficiency, but complex and data-hungry.

Control frequency race: RT-2 and OpenVLA run ~10Hz (100ms+ latency); π0 reaches 50Hz (20ms); Figure's Helix claims 200Hz+ (<5ms). Each order-of-magnitude increase brings exponential engineering complexity.

The Business Anatomy: Charity or Calculation?

Four open-source motivations:

1. Ecosystem lock-in (NVIDIA): free model → developer influx → dependency on chips/simulation → the shovel-seller wins (the CUDA playbook). 2. Data flywheel (Google): open deployment → real-world data → stronger Gemini → better robot models. 3. Standards-setting (academia): technical leadership becomes de facto standard—OpenVLA is already the open robot-model base. 4. Hardware traffic (Unitree): free software sells hardware, like Android.

True vs. fake open source: genuine projects (OpenVLA, Octo, π0 via openpi) release full training code, weights, and permissive licenses (Apache/MIT). Suspicious "open source" releases expose only inference code, restrict weights, or hide data provenance.

Closed-source moats remain: Tesla's millions of vehicles collecting real-world data, Figure's Helix plus in-house humanoid hardware, Boston Dynamics' decade of motion control. Open source may match models, but trails on data and hardware.

Can Open Source Beat the Giants?

Open source's play is a three-layer combination—models (OpenVLA, Octo, π0), data (Open X-Embodiment, DROID), and tools (Isaac Sim, LeRobot)—creating a flywheel: more developers → more data → better models → more developers. Individually, it loses to Google on talent/compute, Tesla on data, and NVIDIA on tooling.

Why now: VLA paradigm is newly established with no monopoly; open datasets and cloud compute democratize access; talent flows to startups. But the window is closing as Tesla's Optimus scales in factories.

Three possible outcomes: open-source victory (30%), differentiated coexistence—open for research/education, closed for commercial mass production (50%), or closed-source monopoly (20%).

Future Directions

From seeing to sensing: adding tactile, auditory, and proprioceptive inputs.
From imitation to understanding: physical world models with intuition for gravity, friction, and inertia.
From single-task to lifelong learning: online adaptation and cross-task transfer.

The ultimate question—who defines the robot brain?—is one of power, not just technology. The most likely result is a hybrid: government-regulated safety layers, open-source-driven application innovation, and commercial competition at the hardware level.

Core Resource Quick Reference

| Model | Link | Params | License | |---|---|---|---| | OpenVLA | github.com/openvla/openvla | 7B | Apache 2.0 | | Octo | github.com/octo-models/octo | 27M–93M | MIT | | π0/π0.5 | github.com/Physical-Intelligence/openpi | 5B+ | Apache 2.0 | | GR00T N1 | developer.nvidia.com/gr00t | undisclosed | NVIDIA license | | Xiaomi-Robotics-0 | github.com/XiaomiRoboticsLab | undisclosed | unclear | | LeRobot | github.com/huggingface/lerobot | - | Apache 2.0 |

Further reading:

OpenVLA paper: arxiv.org/abs/2406.09246
π0.5 paper: arxiv.org/abs/2504.16054
Octo paper: arxiv.org/abs/2405.12213
GR00T N1 technical report: developer.nvidia.com/gr00t

Open source is not the destination—it is the starting point of a movement to democratize robotics technology, and the answer to who controls robot minds will shape the intelligent world we live in.