TikTok parent company ByteDance has demonstrated a robotics and AI system that it said could theoretically perform daily household tasks based on users’ prompts.
The company said its GR-3 is a large-scale vision-language-action model could be paired with its ByteMini, a laboratory-stage robot with two manipulator arms, to carry out simple tasks.
In a video, it demonstrated the robot inserting a hanger into a shirt and hanging it up.
Natural-language prompts
It said the GR-3 model could enable robots to follow natural language instructions and carry out tasks on previously unseen items, in new environments, while dealing with abstract concepts and spatial relationships.
The project is the latest example of a large tech company seeking to apply AI to spatial applications, amidst an ongoing boom in artificial intelligence investment.
In a technical report, ByteDance, which has invested heavily in AI and has attained popularity in China and abroad with its own AI chatbot, said the GR-3 model enabled a robot to follow instructions to pick up an individual item from amongst others and place it in a designated spot.
It could identify an item from its name, size, or spatial relationships, such as “the larger item” or “the item on the left”.
It could follow a single prompt such as “cleaning up the dining table” to complete an entire complex task.
ByteDance, based in Beijing, said it trained GR-3 with massive vision-language data, fine-tuning from human trajectory data collected via virtual reality devices, and imitation learning with robot trajectory data.
“We hope GR-3 can serve as a step towards building generalist robots capable of assisting humans in daily life,” the company’s robotics team said.
AI investment
The initiative comes from ByteDance’s Seed department, which handles its AI and large language model development and was established in 2023 following the successful launch of OpenAI’s ChatGPT.
In March researchers from ByteDance, Tsinghua University and the University of Hong Kong released an open-source system for AI reinforcement learning that they said outperformed a reasoning system from start-up DeepSeek.
The DAPO (Dynamic Sampling Policy Optimisation) system is designed to provide reinforcement-learning techniques for large language models (LLMs) that can be reused by other researchers.
ByteDance’s Doubao chatbot has become China’s most popular chatbot since its launch in May 2024, ranking as the world’s second most popular after OpenAI’s ChatGPT.
