Home Technology peripherals AI Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models

Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models

Jun 09, 2024 pm 05:06 PM
industry

iVideoGPT meets the high interactivity needs of the world model.

Generative models have made significant progress in recent years, among which video generation is becoming a new frontier. An important application of these generative video models is to learn in an unsupervised manner on diverse Internet-scale data for building predictive world models. These world models are expected to accumulate common-sense knowledge about how the world works, allowing predictions of potential future outcomes based on the behavior of agents.

By leveraging these world models, reinforcement learning-based agents can imagine, reason, and plan within the world model, allowing them to perform tasks in the real world with just a few experiments. Acquire new skills more safely and efficiently.

Despite the fundamental connection between generative models and world models, there are still significant differences between the development of generative models for video generation and world models for agent learning. the gap. One of the main challenges is how to achieve the best balance between interactivity and scalability.

#In the field of model-based reinforcement learning, world models mainly use recurrent network architecture. This design facilitates interactive behavioral learning by allowing observations or latent states to be passed based on actions at each step. However, these models mostly focus on game or simulation environments, have simple data, and have limited ability to model large-scale and complex in-the-wild data.

In contrast, Internet-scale video generation models can synthesize realistic long videos that can be controlled with textual descriptions or future action sequences. While such models allow for high-level, long-term planning, their trajectory-level interactivity does not provide agents with sufficient granularity to effectively learn precise behaviors as a fundamental skill.

Researchers from Tsinghua University, Huawei Noah's Ark Laboratory, and Tianjin University proposed iVideoGPT (Interactive VideoGPT), which is an scalable autoregressive Transformer framework. It integrates multi-modal signals (visual observations, actions and rewards) into a series of tokens, enabling the agent to conduct interactive experiences by predicting the next token.

iVideoGPT uses novel compression tokenization technology to effectively discrete high-dimensional visual observations. Leveraging its scalable architecture, researchers were able to pre-train iVideoGPT on millions of human and robot operation trajectories, thereby establishing a versatile foundation that can be used as an interactive world model for a variety of downstream tasks. This research promotes the development of interactive universal world models.
Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models
  • Paper address: https://arxiv.org/pdf/2405.15223
  • Paper title: iVideoGPT: Interactive VideoGPTs are Scalable World Models

Method

In this part, the research team introduces a scalable world model architecture - iVideoGPT, which is extremely flexible and can integrate multi-modal information, including visual observations, actions, rewards and other potential inputs.

#The core of iVideoGPT includes a compression tokenizer for discretizing video frames, and an autoregressive transformer for predicting subsequent tokens. By pre-training on diverse video data, the model can acquire extensive world knowledge and then efficiently transfer to downstream tasks.
Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models
Architecture

Compression tokenization. Transformer performs particularly well at handling sequences of discrete tokens. VQGAN is a commonly used visual tokenizer that converts raw pixels into discrete tokens. The researchers proposed to use a new conditional VQGAN consisting of a dual encoder and decoder {(E_c, D_c), (E_p, D_p)} to tokenize the video.

As shown in Figure 3a, the initial context frame Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models contains rich context information and is tokenized and reconstructed independently through N tokens:

Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models

In contrast, due to the temporal redundancy between context frames and future frames, only necessary change information, such as the position and pose of moving objects, needs to be encoded. The above process is achieved by using conditional encoders and decoders:

Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models

#The researchers implemented the conditional mechanism by using cross-attention between multi-scale feature maps. In general, the tokenizer is trained with the following goals:

Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models

The tokenization proposed in this study mainly has two benefits:

  • First, it significantly reduces the sequence length of the tokenized video, which grows linearly with the number of frames, but the growth rate n is much smaller;
  • Secondly, through conditional encoding, the transformer that predicts subsequent tokens can more easily maintain the temporal consistency of the context and focus on modeling the necessary dynamic information.

Interactive predictions for Transformer. After tokenization, the video is flattened into a series of tokens:

Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models

length is Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models. Special slot tokens [S] are inserted to delineate frame boundaries and facilitate the fusion of additional low-dimensional modalities such as actions. As shown in Figure 3b, a GPT-like autoregressive transformer is used for interactive video prediction by generating next-tokens frame by frame. In this work, the team used the model size of GPT-2 but adapted the LLaMA architecture in order to take advantage of recent innovations in LLM architectures, such as rotational position embedding.

Pre-training

Large language models can be self-supervised through next-word prediction way to gain extensive knowledge from Internet texts. Similarly, the action-free video pre-training paradigm of the world model uses video prediction as the pre-training goal to provide Internet-scale supervision for the physical world knowledge that LLM lacks.

The researchers pre-trained iVideoGPT on this general goal, applying cross-entropy loss to predict subsequent video tokens:

Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models

Pre-training data. Although there are a large number of videos available on the Internet, due to computational limitations, the researchers pretrained iVideoGPT specifically for the field of robotic manipulation. They utilized a mix of 35 datasets from the Open X-Embodiment (OXE) dataset and the Something-Something v2 (SSv2) dataset, totaling 1.5 million trajectories.

Fine-tuning

action conditions and reward prediction. The team’s architecture is designed to flexibly integrate additional modalities to learn an interactive world model, as shown in Figure 3b. Actions are integrated via linear projection and added to slot token embeddings. For reward prediction, instead of learning a separate reward predictor, they added a linear head on the hidden state of the last token of each observation.

This multi-task learning method can enhance the model's attention to task-related information, thereby improving the prediction accuracy of control tasks. In addition to the cross-entropy loss of Equation (3), they also used the mean square error loss for reward prediction.

Tokenizer Adapt. The research team chose to update the complete model, including the tokenizer, to adapt to downstream tasks, and found this strategy to be more effective than parameter-efficient fine-tuning methods.

#There is little literature exploring the use of VQGAN tokenizer for domain-specific data. In this work, since tokenization decouples dynamic information from contextual conditions, it is assumed that although this model may encounter unseen objects in downstream tasks, such as different types of robots, the transformer learns from diverse scenarios Basic physics knowledge - such as movement and interaction - is shared.

This hypothesis is supported by experiments where they migrated iVideoGPT from mixed pre-trained data to the unseen BAIR dataset, where the pre-trained transformer can zero-sample Generalizing to predict natural motion requires only fine-tuning the tokenizer for unseen robot grippers (see Figure 7). This feature is particularly important for scaling GPT-like transformers to large sizes, enabling lightweight alignment across domains while keeping the transformer intact.
Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models
Experiment

As shown in Table 1, compared with the SOTA method, iVideoGPT Demonstrates competitive performance while enabling interactivity and scalability in its architecture. Although preliminary experiments were performed at the low resolution of 64×64, iVideoGPT can be easily extended to RoboNet’s 256×256.
Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models
See Figure 9 for qualitative results.
Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models
Figure 4 shows the success rate of iVideoGPT compared to the baseline model. iVideoGPT significantly outperforms all baselines on both RoboDesk tasks and achieves average performance comparable to the strongest model SVG'.
Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models
Figure 6 shows that the model-based algorithm not only improves the sample efficiency than the model-free algorithm, but also meets or exceeds the performance of DreamerV3.
Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models
The next study analyzes the zero-shot video prediction capabilities of large-scale pre-trained iVideoGPT on the unseen BAIR dataset. Interestingly, we observe in the second row of Figure 7 that iVideoGPT predicts the natural motion of a robot gripper without fine-tuning—albeit differently from the pre-training dataset. This shows that although the model has limited zero-shot generalization to completely unseen robots due to insufficient diversity in the pre-training data, it effectively separates scene context from motion dynamics. In contrast, using the adapted tokenizer, the non-fine-tuned Transformer successfully transfers the pre-trained knowledge and predicts the motion of the novel robot in the third row, providing similar perception to the fully fine-tuned Transformer in the fourth row. The quality and quantitative results are shown in Figure 8a.
Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models
For more results, please refer to the original paper.

The above is the detailed content of Tsinghua University, Huawei and others proposed iVideoGPT: specializing in interactive world models. For more information, please follow other related articles on the PHP Chinese website!

Statement of this Website
The content of this article is voluntarily contributed by netizens, and the copyright belongs to the original author. This site does not assume corresponding legal responsibility. If you find any content suspected of plagiarism or infringement, please contact admin@php.cn

Hot AI Tools

Undresser.AI Undress

Undresser.AI Undress

AI-powered app for creating realistic nude photos

AI Clothes Remover

AI Clothes Remover

Online AI tool for removing clothes from photos.

Undress AI Tool

Undress AI Tool

Undress images for free

Clothoff.io

Clothoff.io

AI clothes remover

AI Hentai Generator

AI Hentai Generator

Generate AI Hentai for free.

Hot Article

R.E.P.O. Energy Crystals Explained and What They Do (Yellow Crystal)
1 months ago By 尊渡假赌尊渡假赌尊渡假赌
R.E.P.O. Best Graphic Settings
1 months ago By 尊渡假赌尊渡假赌尊渡假赌
R.E.P.O. How to Fix Audio if You Can't Hear Anyone
1 months ago By 尊渡假赌尊渡假赌尊渡假赌
R.E.P.O. Chat Commands and How to Use Them
1 months ago By 尊渡假赌尊渡假赌尊渡假赌

Hot Tools

Notepad++7.3.1

Notepad++7.3.1

Easy-to-use and free code editor

SublimeText3 Chinese version

SublimeText3 Chinese version

Chinese version, very easy to use

Zend Studio 13.0.1

Zend Studio 13.0.1

Powerful PHP integrated development environment

Dreamweaver CS6

Dreamweaver CS6

Visual web development tools

SublimeText3 Mac version

SublimeText3 Mac version

God-level code editing software (SublimeText3)

DeepMind robot plays table tennis, and its forehand and backhand slip into the air, completely defeating human beginners DeepMind robot plays table tennis, and its forehand and backhand slip into the air, completely defeating human beginners Aug 09, 2024 pm 04:01 PM

But maybe he can’t defeat the old man in the park? The Paris Olympic Games are in full swing, and table tennis has attracted much attention. At the same time, robots have also made new breakthroughs in playing table tennis. Just now, DeepMind proposed the first learning robot agent that can reach the level of human amateur players in competitive table tennis. Paper address: https://arxiv.org/pdf/2408.03906 How good is the DeepMind robot at playing table tennis? Probably on par with human amateur players: both forehand and backhand: the opponent uses a variety of playing styles, and the robot can also withstand: receiving serves with different spins: However, the intensity of the game does not seem to be as intense as the old man in the park. For robots, table tennis

The first mechanical claw! Yuanluobao appeared at the 2024 World Robot Conference and released the first chess robot that can enter the home The first mechanical claw! Yuanluobao appeared at the 2024 World Robot Conference and released the first chess robot that can enter the home Aug 21, 2024 pm 07:33 PM

On August 21, the 2024 World Robot Conference was grandly held in Beijing. SenseTime's home robot brand "Yuanluobot SenseRobot" has unveiled its entire family of products, and recently released the Yuanluobot AI chess-playing robot - Chess Professional Edition (hereinafter referred to as "Yuanluobot SenseRobot"), becoming the world's first A chess robot for the home. As the third chess-playing robot product of Yuanluobo, the new Guoxiang robot has undergone a large number of special technical upgrades and innovations in AI and engineering machinery. For the first time, it has realized the ability to pick up three-dimensional chess pieces through mechanical claws on a home robot, and perform human-machine Functions such as chess playing, everyone playing chess, notation review, etc.

Claude has become lazy too! Netizen: Learn to give yourself a holiday Claude has become lazy too! Netizen: Learn to give yourself a holiday Sep 02, 2024 pm 01:56 PM

The start of school is about to begin, and it’s not just the students who are about to start the new semester who should take care of themselves, but also the large AI models. Some time ago, Reddit was filled with netizens complaining that Claude was getting lazy. "Its level has dropped a lot, it often pauses, and even the output becomes very short. In the first week of release, it could translate a full 4-page document at once, but now it can't even output half a page!" https:// www.reddit.com/r/ClaudeAI/comments/1by8rw8/something_just_feels_wrong_with_claude_in_the/ in a post titled "Totally disappointed with Claude", full of

At the World Robot Conference, this domestic robot carrying 'the hope of future elderly care' was surrounded At the World Robot Conference, this domestic robot carrying 'the hope of future elderly care' was surrounded Aug 22, 2024 pm 10:35 PM

At the World Robot Conference being held in Beijing, the display of humanoid robots has become the absolute focus of the scene. At the Stardust Intelligent booth, the AI ​​robot assistant S1 performed three major performances of dulcimer, martial arts, and calligraphy in one exhibition area, capable of both literary and martial arts. , attracted a large number of professional audiences and media. The elegant playing on the elastic strings allows the S1 to demonstrate fine operation and absolute control with speed, strength and precision. CCTV News conducted a special report on the imitation learning and intelligent control behind "Calligraphy". Company founder Lai Jie explained that behind the silky movements, the hardware side pursues the best force control and the most human-like body indicators (speed, load) etc.), but on the AI ​​side, the real movement data of people is collected, allowing the robot to become stronger when it encounters a strong situation and learn to evolve quickly. And agile

ACL 2024 Awards Announced: One of the Best Papers on Oracle Deciphering by HuaTech, GloVe Time Test Award ACL 2024 Awards Announced: One of the Best Papers on Oracle Deciphering by HuaTech, GloVe Time Test Award Aug 15, 2024 pm 04:37 PM

At this ACL conference, contributors have gained a lot. The six-day ACL2024 is being held in Bangkok, Thailand. ACL is the top international conference in the field of computational linguistics and natural language processing. It is organized by the International Association for Computational Linguistics and is held annually. ACL has always ranked first in academic influence in the field of NLP, and it is also a CCF-A recommended conference. This year's ACL conference is the 62nd and has received more than 400 cutting-edge works in the field of NLP. Yesterday afternoon, the conference announced the best paper and other awards. This time, there are 7 Best Paper Awards (two unpublished), 1 Best Theme Paper Award, and 35 Outstanding Paper Awards. The conference also awarded 3 Resource Paper Awards (ResourceAward) and Social Impact Award (

Hongmeng Smart Travel S9 and full-scenario new product launch conference, a number of blockbuster new products were released together Hongmeng Smart Travel S9 and full-scenario new product launch conference, a number of blockbuster new products were released together Aug 08, 2024 am 07:02 AM

This afternoon, Hongmeng Zhixing officially welcomed new brands and new cars. On August 6, Huawei held the Hongmeng Smart Xingxing S9 and Huawei full-scenario new product launch conference, bringing the panoramic smart flagship sedan Xiangjie S9, the new M7Pro and Huawei novaFlip, MatePad Pro 12.2 inches, the new MatePad Air, Huawei Bisheng With many new all-scenario smart products including the laser printer X1 series, FreeBuds6i, WATCHFIT3 and smart screen S5Pro, from smart travel, smart office to smart wear, Huawei continues to build a full-scenario smart ecosystem to bring consumers a smart experience of the Internet of Everything. Hongmeng Zhixing: In-depth empowerment to promote the upgrading of the smart car industry Huawei joins hands with Chinese automotive industry partners to provide

Li Feifei's team proposed ReKep to give robots spatial intelligence and integrate GPT-4o Li Feifei's team proposed ReKep to give robots spatial intelligence and integrate GPT-4o Sep 03, 2024 pm 05:18 PM

Deep integration of vision and robot learning. When two robot hands work together smoothly to fold clothes, pour tea, and pack shoes, coupled with the 1X humanoid robot NEO that has been making headlines recently, you may have a feeling: we seem to be entering the age of robots. In fact, these silky movements are the product of advanced robotic technology + exquisite frame design + multi-modal large models. We know that useful robots often require complex and exquisite interactions with the environment, and the environment can be represented as constraints in the spatial and temporal domains. For example, if you want a robot to pour tea, the robot first needs to grasp the handle of the teapot and keep it upright without spilling the tea, then move it smoothly until the mouth of the pot is aligned with the mouth of the cup, and then tilt the teapot at a certain angle. . this

Distributed Artificial Intelligence Conference DAI 2024 Call for Papers: Agent Day, Richard Sutton, the father of reinforcement learning, will attend! Yan Shuicheng, Sergey Levine and DeepMind scientists will give keynote speeches Distributed Artificial Intelligence Conference DAI 2024 Call for Papers: Agent Day, Richard Sutton, the father of reinforcement learning, will attend! Yan Shuicheng, Sergey Levine and DeepMind scientists will give keynote speeches Aug 22, 2024 pm 08:02 PM

Conference Introduction With the rapid development of science and technology, artificial intelligence has become an important force in promoting social progress. In this era, we are fortunate to witness and participate in the innovation and application of Distributed Artificial Intelligence (DAI). Distributed artificial intelligence is an important branch of the field of artificial intelligence, which has attracted more and more attention in recent years. Agents based on large language models (LLM) have suddenly emerged. By combining the powerful language understanding and generation capabilities of large models, they have shown great potential in natural language interaction, knowledge reasoning, task planning, etc. AIAgent is taking over the big language model and has become a hot topic in the current AI circle. Au

See all articles