EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·Sep 6·5 min read
GPT-6 Astra on robot arms
TL;DR
OpenAI's GPT-6 Astra significantly outperformed previous models in controlling robot arms for a block-picking task, achieving 19 out of 20 successful trials. However, it struggled with a puzzle piece insertion task, completing only 2 out of 20 attempts.
✦ Why It Matters
Consider testing GPT-6 Astra for your robotic applications, especially for simple manipulation tasks.
Key Takeaways
- Utilize GPT-6 Astra for high success rates in simple robotic tasks.
- Evaluate cost-effectiveness when choosing AI models for robotic applications.
- Be cautious of Astra's limitations in complex manipulation tasks.
- Consider the specific rig setup when comparing model performances.
Full Summary
In a recent evaluation, OpenAI's GPT-6 Astra was tested against Claude Fable 5 and Fable 5.1 on two robotic manipulation tasks using YAM arms. For the task of placing a red block into a bowl, Astra succeeded in 19 out of 20 trials, achieving a mean score of 3.95 and costing approximately $0.94 per run.
In contrast, Fable 5.1 completed 8 out of 20 trials at a higher cost of $2.12. However, Astra's performance on the puzzle task was less impressive, with only 2 completions out of 20, mirroring Fable 5.1's results.
The trials were conducted under a controlled policy with specific observation and control parameters. These findings highlight Astra's potential for cost-effective robotic manipulation, though it also reveals limitations in more complex tasks.