RobotGuide
Brain and cerebellum/Model survey

A map of the brain layer

Models worth remembering

Skip the news cycle; memorize routes instead. The first seven models cover the four main VLA approaches; the last four fill in the gaps. What each one actually solved is more useful than its parameter count.

RT-2

Google DeepMind · 2023

The pioneering VLA Closed-source 1–3Hz 5B / 55B (per official papers)

In one sentence: first proof that robots can understand human language — even commands they were never taught.

First proof that wiring a vision-language model directly to action outputs lets robots follow unseen instructions like "bring me the dinosaur toy" — internet knowledge transfers straight to robots. It established the VLA paradigm; every later model builds on its skeleton.

2026 status: Superseded by successors: its paradigm legacy lives in the RT-X dataset and the skeleton of all later models; Google's large-model line has advanced to Gemini Robotics 2.

Inclusion criteria: models with verifiable impact on industry roadmaps. 2026-11 update: each card now includes a '2026 status' field, and four gap-filling models were added (engineering convergence, lightweight, dual-arm, and data-driven, one each).

OpenVLA

Stanford / Berkeley et al. · 2024

Open-source VLA Open weights About 3–6Hz (per paper) 7B

In one sentence: an open-source recreation of RT-2 that fine-tunes on a single GPU — where students start with VLAs.

An open-source VLA trained on the Open X-Embodiment cross-embodiment dataset, matching the closed-source RT-2-X in performance. What matters most for students: it can be fine-tuned on a single GPU — the standard starting point for learning VLA training.

2026 status: The community mainline is now OpenVLA-OFT: FAST frequency-domain tokenization + a flow-matching head; still the smoothest fine-tuning starting point.

Inclusion criteria: models with verifiable impact on industry roadmaps. 2026-11 update: each card now includes a '2026 status' field, and four gap-filling models were added (engineering convergence, lightweight, dual-arm, and data-driven, one each).

π0 (pi-zero)

Physical Intelligence · 2024

VLA · action-expert route Open-sourced (openpi) 50Hz action output (official figures) 3B-class base + action expert

In one sentence: solves the fast-and-smooth action problem — continuous actions are no longer split into individual tokens.

The core innovation is in the action head: it uses flow matching to generate continuous action streams instead of discretizing actions into tokens. High-frequency continuous control plus long-horizon dexterous tasks like folding clothes are its signature capabilities. The open-source openpi lets teams reproduce it directly.

2026 status: π0.5 adds open-world generalization; π0.7 adds a world-model module (generating subgoal images); since 2026-10 PI has been beta-testing a hosted API. The open-source openpi remains one of the preferred fine-tuning bases.

Inclusion criteria: models with verifiable impact on industry roadmaps. 2026-11 update: each card now includes a '2026 status' field, and four gap-filling models were added (engineering convergence, lightweight, dual-arm, and data-driven, one each).

Helix

Figure · 2025

Dual-system VLA (brain and cerebellum) Closed-source S2 about 7–9Hz + S1 200Hz (per official blog) S2 about 7B + S1 about 80M

In one sentence: brain and cerebellum packed separately into the robot itself — cerebellum at 200Hz, brain at 7-9Hz.

The representative case of shipping the 'brain S2 + cerebellum S1' split as a product running on the robot's own GPU. The cerebellum handles only 200Hz continuous control of the upper body while the brain handles understanding and intent — the two-layer interface design (a single intent vector) is its most valuable engineering decision.

2026 status: Still closed-source; the dual-system architecture is widely referenced by GR00T, Xiaomi Robotics-0 and others, while Figure keeps iterating its hardware line.

Inclusion criteria: models with verifiable impact on industry roadmaps. 2026-11 update: each card now includes a '2026 status' field, and four gap-filling models were added (engineering convergence, lightweight, dual-arm, and data-driven, one each).

GR00T N1

NVIDIA · 2025

Open-source humanoid base + simulation data pipeline Open weights Dual-system architecture (official figures) VLM backbone + DiT diffusion action head

In one sentence: NVIDIA doesn't sell a model; it sells the whole development stack of 'model + simulation data pipeline'.

More than a model: it ships with an Isaac-based synthetic-data simulation pipeline (GR00T-Blueprint) that tries to solve the fundamental VLA bottleneck of expensive real-robot data. A platform play — what it sells is the whole development stack.

2026 status: Now at N1.7 (2026-04 early access); the N1.6 backbone (Cosmos-2B variant vs Eagle) has inconsistent official accounts; N2 is teased for late 2026; license terms conflict between the repo and blog — check the LICENSE before use.

Inclusion criteria: models with verifiable impact on industry roadmaps. 2026-11 update: each card now includes a '2026 status' field, and four gap-filling models were added (engineering convergence, lightweight, dual-arm, and data-driven, one each).

Gemini Robotics

Google DeepMind · 2025

Multimodal base + action layer Closed-source Based on Gemini 2.0 (official figures)

In one sentence: the base-model giant's route — don't build robots, license the model to those who do.

The robotics extension of a multimodal base model: Gemini Robotics-ER handles 'embodied reasoning' while the action layer executes. The base-model vendor's general route — build no hardware, license the model to robot makers.

2026 status: On 2026-07-30 Google shipped the Gemini Robotics 2 trio (VLA + ER 2 embodied reasoning + On-Device 2) with whole-body control and multi-robot collaboration; the '<200 samples to adapt to a new robot' figure comes from this release.

Inclusion criteria: models with verifiable impact on industry roadmaps. 2026-11 update: each card now includes a '2026 status' field, and four gap-filling models were added (engineering convergence, lightweight, dual-arm, and data-driven, one each).

GO-1

Zhiyuan Robotics (AgiBot) · 2025

Chinese VLA · data companion Model and AgiBot World dataset open-sourced ViLLA architecture (VLM + latent action model + action expert)

In one sentence: China's representative, shipping the largest open-sourced Chinese real-robot dataset to date.

China's representative VLA: the ViLLA architecture adds an action expert on top of an open-source VLM, paired with the open-source million-scale real-robot dataset AgiBot World. The dataset's open-source value may exceed the model's own.

2026 status: AgiBot World keeps expanding (million-scale real-robot trajectories open-sourced); the data advantage remains the core of this line, and the ViLLA architecture is being reused by the open-source community.

Inclusion criteria: models with verifiable impact on industry roadmaps. 2026-11 update: each card now includes a '2026 status' field, and four gap-filling models were added (engineering convergence, lightweight, dual-arm, and data-driven, one each).

OpenVLA-OFT

Stanford / Berkeley et al. · 2025

Open-source fine-tuning starting point Open-source (MIT/LLAMA2 mixed license) ~10-30Hz 7B

In one sentence: a speed-optimized rework of OpenVLA — FAST compression + flow-matching head, the first stop for open-source fine-tuning.

Replaces OpenVLA's discrete-token action head with flow matching and compresses sequences with FAST frequency-domain compression, improving both inference efficiency and the force-control ceiling. It represents no new paradigm — it represents the engineering convergence of the open-source baseline.

2026 status: One of the default starting points in the open-source fine-tuning ecosystem.

Inclusion criteria: models with verifiable impact on industry roadmaps. 2026-11 update: each card now includes a '2026 status' field, and four gap-filling models were added (engineering convergence, lightweight, dual-arm, and data-driven, one each).

SmolVLA

Hugging Face · 2025

Lightweight VLA Open-source (Apache-2.0) Asynchronous inference 450M

In one sentence: 450M parameters running on consumer hardware, community data + the LeRobot workflow.

Proof that VLAs can be small: trained on community open data, runs on consumer hardware, asynchronous inference built in. The best starting point for teaching and prototype validation.

2026 status: The LeRobot ecosystem's default teaching model.

Inclusion criteria: models with verifiable impact on industry roadmaps. 2026-11 update: each card now includes a '2026 status' field, and four gap-filling models were added (engineering convergence, lightweight, dual-arm, and data-driven, one each).

RDT-1B

Tsinghua University · 2025

Diffusion action head · dual-arm Open-source 1B

In one sentence: a 1B diffusion Transformer, the representative Chinese open-source model for dual-arm manipulation.

Uses a diffusion Transformer to produce dual-arm continuous actions and is widely used as a baseline on dual-arm tasks. Within the Chinese open-source line it serves as the counterpart to the π0 route.

2026 status: A common baseline for dual-arm tasks.

Inclusion criteria: models with verifiable impact on industry roadmaps. 2026-11 update: each card now includes a '2026 status' field, and four gap-filling models were added (engineering convergence, lightweight, dual-arm, and data-driven, one each).

Xiaomi Robotics-1

Xiaomi · 2026

Real-robot data-driven base Not open-sourced 80ms (Robotics-0 consumer-grade)

In one sentence: pre-trained on 100,000 hours of real-robot data; dual simulation leaderboard #1 in 2026-07.

A mixed brain-and-cerebellum architecture (Robotics-0 runs 80ms inference on consumer GPUs), with the base pre-trained on 100,000 hours of real-robot manipulation trajectories — currently the strongest example of the 'data volume beats architecture tricks' route.

2026 status: Dual #1 on RoboCasa 365 / RoboDojo in 2026-07; validated 8 continuous hours on a real factory floor.

Inclusion criteria: models with verifiable impact on industry roadmaps. 2026-11 update: each card now includes a '2026 status' field, and four gap-filling models were added (engineering convergence, lightweight, dual-arm, and data-driven, one each).

First move if you want into this direction

Not reading papers — reproducing: fine-tune OpenVLA weights once on a single GPU, or run one trajectory with openpi (the pi0 open release), or train a small action policy on the AgiBot World open dataset. One complete «data → training → evaluation on a real robot or in sim» beats ten close readings. Matching topics: C8, F1, F4.

This field turns over every three months

This page only lists models with verifiable impact on industry routes. To keep up: watch three repos (OpenVLA, openpi, GR00T) and two conferences (RSS, CoRL). Skip the keynote clips on social media.

Keep reading

Brain and cerebellum: the full principle → Six company teardowns → The matching roles →