Robbyant open-sources LingBot-VLA 2.0 robot foundation model post-training code

The model beats rival systems on bimanual tabletop tasks.

Combined views

16.1K

11 posts, first seen 6h ago

127 likes

Commentary on X

GU

Gaurav Upreti@gauptx

@omarsar0 Training a policy that generalizes across multiple robot embodiments is arguably a harder problem than optimizing for one platform. Generalization—not specialization—is what will make robotics scalable.

From the lingbot paper. Such a cool way of demonstrating robot actions.

6:26 AM · Jul 15, 2026

Alina Fomina@fominaaalina

AF

@omarsar0 need it in my kitchen

8:02 AM · Jul 15, 2026

LingBot-VLA 2.0 is an impressive new embodied model. Open source and is trained across diverse robot configurations, from single-arm robots to humanoid platforms. It packs 60K hours of curated robot and human data into one generalist policy. It improves robots on difficult long-horizon tasks. Great release by @robbyant_brain.

Combined views

16.1K

11 posts, first seen 6h ago

127 likes16 comments

16 comments

93 saves

14 reposts

Sentiment

Positive100%0%Negative

Positive Read

Users are excited about LingBot-VA 2.0 and the open-source LingBot-VLA 2.0 embodied model because of its cool demonstrations of robot actions, strong generalization across embodiments, and practical potential like kitchen use.

Based on 3 visible X reactions from 5 accounts; directional sample.

Useful Links

Digg Deeper

Ask a question below.

Published answers will appear here.

Commentary on X

GU

Gaurav Upreti@gauptx

@omarsar0 Training a policy that generalizes across multiple robot embodiments is arguably a harder problem than optimizing for one platform. Generalization—not specialization—is what will make robotics scalable.

From the lingbot paper. Such a cool way of demonstrating robot actions.

6:26 AM · Jul 15, 2026

Alina Fomina@fominaaalina

AF

@omarsar0 need it in my kitchen

8:02 AM · Jul 15, 2026

LingBot-VLA 2.0 is an impressive new embodied model. Open source and is trained across diverse robot configurations, from single-arm robots to humanoid platforms. It packs 60K hours of curated robot and human data into one generalist policy. It improves robots on difficult long-horizon tasks. Great release by @robbyant_brain.

View all

Let’s talk about the data first. In robotics, quality data is a big challenge. Robbyant curated 60,000 hours of pretraining data. That covers 50,000 hours of real-robot trajectories and 10,000 hours of egocentric human video. Robbyant filters the raw data aggressively to hold quality high.

LingBot-VLA 2.0 shows practicality too. Robbyant open-sourced the post-training code. Inference takes about 130 ms on a single NVIDIA GeForce RTX 4090D using 10 denoising steps. That means you can adapt and test it without a cluster.

LingBot-VLA 2.0 also learns to predict what the scene will look like next. The model predicts future depth and semantic features using LingBot-Depth and a causal video model. So it has a sense of how the task will unfold before it acts.

View all

Let’s talk about the data first. In robotics, quality data is a big challenge. Robbyant curated 60,000 hours of pretraining data. That covers 50,000 hours of real-robot trajectories and 10,000 hours of egocentric human video. Robbyant filters the raw data aggressively to hold quality high.

LingBot-VLA 2.0 shows practicality too. Robbyant open-sourced the post-training code. Inference takes about 130 ms on a single NVIDIA GeForce RTX 4090D using 10 denoising steps. That means you can adapt and test it without a cluster.

LingBot-VLA 2.0 also learns to predict what the scene will look like next. The model predicts future depth and semantic features using LingBot-Depth and a causal video model. So it has a sense of how the task will unfold before it acts.

Future prediction pays off once manipulation leaves the tabletop. The model coordinates its head, waist, mobile base, arms, and hands across long tasks, such as navigating to an appliance and opening it. One policy covers different robot bodies, long tasks, and movement around the room.

The embodiment coverage is wide. Pretraining spans 20 configurations, from single-arm robots and dual-arm rigs to full humanoids like Unitree G1 and Fourier GR-2. VLA models can work across all of them.

The action space reaches past arms and grippers. Dexterous hands, head, waist, and mobile base all map into one representation. They show that the same policy interface works across different hardware.

Their in-context learning pipeline is very cool

6:24 AM · Jul 15, 2026

93 saves

14 reposts

Sentiment

Positive100%0%Negative

Positive Read

Users are excited about LingBot-VA 2.0 and the open-source LingBot-VLA 2.0 embodied model because of its cool demonstrations of robot actions, strong generalization across embodiments, and practical potential like kitchen use.

Based on 3 visible X reactions from 5 accounts; directional sample.

Useful Links

Digg Deeper

Ask a question below.

Published answers will appear here.

Future prediction pays off once manipulation leaves the tabletop. The model coordinates its head, waist, mobile base, arms, and hands across long tasks, such as navigating to an appliance and opening it. One policy covers different robot bodies, long tasks, and movement around the room.

The embodiment coverage is wide. Pretraining spans 20 configurations, from single-arm robots and dual-arm rigs to full humanoids like Unitree G1 and Fourier GR-2. VLA models can work across all of them.

The action space reaches past arms and grippers. Dexterous hands, head, waist, and mobile base all map into one representation. They show that the same policy interface works across different hardware.

Their in-context learning pipeline is very cool

6:24 AM · Jul 15, 2026