OpenAI’s Astra agent successfully controlled a humanoid robot for manipulation tasks without any specific training, using a zero-shot approach that leverages only its general training. During evaluations conducted by third-party organization RoboCurve, the model achieved a 95% success rate, completing 19 out of 20 attempts with I2RT YAM bimanual robot arms. Each run required approximately 2.1k tokens of output and took about 2.5 minutes to complete. In simulated environments, performance was even better with a 98% success rate, completing 49 out of 50 assignments in RoboLab benchmarks. The 3-percentage-point gap between simulation and real-world performance illustrates the challenges posed by unforeseen factors in physical environments.
Source: Read the original article
