Robbyant debuts open-source 6B VLA model for robots with 60k hours of training
Reported by MarkTechPost: Robbyant Releases LingBot-VLA 2.0: An Open-Source 6B Vision-Language-Action (VLA) Model for Cross-Embodiment Robot Manipulation. Analysis and context written by TickrWire.
Robbyant released LingBot-VLA 2.0, an open-source 6-billion-parameter vision-language-action model for robot manipulation across different hardware setups.
- LingBot-VLA 2.0 is an open-source 6B parameter vision-language-action model for robot manipulation.
- Pretrained on 60,000 hours of data, including 50,000 hours of robot trajectories across 20 configurations.
- Uses a token-level Mixture-of-Experts architecture without load-balancing loss for scalability.
- Maps all robot embodiments into a shared 55-dimensional action space for cross-embodiment manipulation.
Robbyant, an Ant Group subsidiary focused on robotics, has launched LingBot-VLA 2.0, an open-source vision-language-action (VLA) model designed to unify robot manipulation across diverse hardware configurations. The 6-billion-parameter model is pretrained on approximately 60,000 hours of data, including 50,000 hours of robot trajectories spanning 20 different robot setups and 10,000 hours of egocentric human video footage.
A key innovation in LingBot-VLA 2.0 is its token-level Mixture-of-Experts architecture, which scales model capacity without requiring a load-balancing loss. The model maps all robot embodiments into a shared 55-dimensional canonical action space, covering a wide range of robotic components such as arms, dexterous hands, waists, heads, and mobile bases. This approach aims to enable seamless cross-embodiment manipulation, allowing robots with different physical configurations to perform tasks using a unified policy.
The release is Apache-2.0 licensed, making it freely available for research and commercial use. Robbyant positions this model as a step toward more generalizable and adaptable robotics systems, potentially reducing the need for task-specific fine-tuning across different robotic platforms.
Provides a powerful open-source tool for building generalizable robot manipulation policies across diverse hardware.
Enables faster development of robotics applications by reducing the need for task-specific fine-tuning.
Highlights Ant Group's strategic push into robotics with a technically advanced open-source model.
Offers a cutting-edge example of combining vision, language, and action in robotics research.
- Vision-Language-Action (VLA) model
- A neural network that integrates visual input, natural language instructions, and action outputs to enable robots to perform tasks based on human commands.
- Mixture-of-Experts (MoE)
- A model architecture where multiple specialized sub-networks (experts) are combined, with only a subset activated per input to improve efficiency and scalability.
- Canonical action space
- A standardized representation of robot actions that unifies diverse hardware configurations into a common framework for manipulation.
RoboticsAmazon aims for delivery drones to reach 500 US neighborhoods by end of 2026
RoboticsI Saw the Future of AI in a Robot That Can Learn on the Spot
ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning
Project Convergence-Capstone 6 validates Artificial Intelligence-Assisted Maintenance in the field - army.mil
Why tactile intelligence is the next layer for physical AI - The World Economic Forum
HardwareCerebras unveils CS-4 with double the performance on the same chip
Cerebras has launched the CS-4, a rack-scale AI accelerator that doubles performance over its predecessor by optimizing power and cooling for the WSE-3 chip.
AI ResearchWho’s behind the new ‘stealth model’ Ox Alpha?
A mysterious reasoning model named Ox Alpha appeared on OpenRouter, prompting widespread speculation regarding its anonymous creator.
SecurityFlock CEO calls for ‘compromise’ as surveillance company faces growing backlash
Flock Safety CEO Garrett Langley is advocating for a compromise between public safety and privacy as the surveillance tech company encounters intense scrutiny and political pushback over alleged misuse.
SecurityIs it legal to train AI models on copyrighted books? It’s complicated
Courts are split on whether training AI on copyrighted books counts as illegal copying or protected fair use, leaving authors and tech firms in legal limbo.
AI ToolsAn AI boss fired its first employee but only after humans reminded it of its own rules
An AI agent running a San Francisco store fired an employee only after humans reminded it of its own termination rules, highlighting gaps in long-term memory and leniency in AI management.
AI ResearchAI could make scientists do more work less well, not less work better, study argues
A theoretical economics study argues that language models might make scientific research shallower because time saved on routine tasks encourages academics to start more projects rather than improve existing ones.