Want to Make Robots Smarter? Tsinghua Team Discovers the Secret to Accelerated Robot Learning

AIbase基地

Published inAI News · 5 min read · Nov 12, 2024

119

The rapid advancement of deep learning is inseparable from the scaling of datasets, models, and computational power. In the fields of natural language processing and computer vision, researchers have discovered a power-law relationship between model performance and data scale. However, in the field of robotics, particularly in robot manipulation, similar scaling laws have not yet been established.

A research team from Tsinghua University recently published a paper exploring the scaling laws in robot imitation learning and proposed an efficient data collection strategy. They were able to collect sufficient data in just an afternoon, enabling the strategy to achieve approximately 90% success rates in new environments and with new objects.

The researchers divided generalization capabilities into environmental generalization and object generalization, and collected human demonstration data using a handheld gripper in various environments and with different objects. They then modeled this data using a diffusion strategy. The researchers initially focused on two tasks: pouring water and placing a mouse. By analyzing how the performance of the strategy varied with the number of training environments or objects, they summarized the scaling laws of data.

The study results indicate:

The generalization capabilities of the strategy to new objects, new environments, or both are related to the number of training objects, training environments, or training environment-object pairs, respectively, following a power-law relationship.

Increasing the diversity of environments and objects is more effective than increasing the number of demonstrations for each environment or object.

Collecting data in as many environments as possible (e.g., 32 environments), with each environment having one unique object and 50 demonstrations, can train a strategy with strong generalization capabilities (90% success rate), enabling it to operate in new environments and with new objects.

Based on these scaling laws, the researchers proposed an efficient data collection strategy. They suggest collecting data in as many different environments as possible, each with a unique object. When the total number of environment-object pairs reaches 32, it is usually sufficient to train a strategy capable of operating in new environments and interacting with previously unseen objects. For each environment-object pair, it is recommended to collect 50 demonstrations.

To verify the general applicability of the data collection strategy, the researchers applied it to two new tasks: folding towels and unplugging chargers. The results showed that the strategy could also train highly generalized strategies for these new tasks.

The study demonstrates that with relatively modest time and resources, it is possible to learn a single-task strategy that can be deployed zero-shot to any environment and object. To further support research in this area, the Tsinghua team has released their code, data, and models, hoping to inspire further research in the field and ultimately achieve a general-purpose robot capable of solving complex, open-world problems.

Paper link: https://arxiv.org/pdf/2410.18647

The Generative AI Market is Expected to Surge to $57.96 Billion by 2028!

Recently, a business research company released its 2024 Generative AI Market Report, showing that the market has rapidly expanded in recent years. According to the report, the generative AI market is expected to grow from $13.64 billion in 2023 to $18.1 billion in 2024, with a compound annual growth rate (CAGR) of 32.7%. This growth is mainly attributed to increased computing power, widespread availability of data, continuous R&D advancements, increased industry applications, and breakthroughs in deep learning. Image Source Note: The image was generated by AI, image usage authorized.

MIT Unveils New Robot Training Model to Tackle Problems in a Simpler, More Direct Way

The Massachusetts Institute of Technology (MIT) showcased a brand new robot training model this week that abandons the previous focus on training methods centered around specific datasets, opting instead for an approach that utilizes vast amounts of information similar to that used in training large language models (LLMs). Researchers noted that imitation learning—where agents learn by mimicking individuals performing tasks—can fail when facing minor challenges. These challenges may include varying lighting conditions, different environmental setups, or new obstacles. In such cases, the robots lack sufficient data to adapt to these changes.

MIT Unveils a Method Inspired by Large Language Models to Teach Robots New Skills for the First Time

This week, MIT showcased a new model for training robots, aimed at addressing the challenges faced by imitation learning when introduced to small obstacles. Researchers pointed out that imitation learning can fail in conditions such as varying lighting, different environments, or new obstacles because robots simply do not have enough data to adapt. The team sought powerful data methods, like those used in models such as GPT-4, to tackle the problem. They introduced a new architecture called the Heterogeneous Pretrained Transformer (HPT), which aggregates information from different sensors and environments.

What Are the Best AI Painting Software? Is Top 1 Midjourney?

AI painting is a revolutionary image generation technology based on deep learning algorithms, particularly Generative Adversarial Networks (GANs) and diffusion models. This approach analyzes vast amounts of image data to learn and simulate human painting techniques, thereby creating entirely new visual works. AI painting not only captures and reproduces the complex details of the real world accurately but also blends different artistic styles, showcasing astonishing creativity and imagination. The core of this technology lies in transforming abstract textual descriptions into concrete visual expressions.