Briefing
In simulation Astra reached a 22.48% average success rate over 2,100 trials with no robot finetuning. GPT-5.5 and DeepSeek-Flash, run through the same interface, scored 0.88% and 1.92%. Astra did well on tasks that need semantic understanding and poorly on precision, dynamic control and two-arm coordination.
The official RoboDojo-Real protocol was never completed. The paper keeps 33 trials on ARX X5, Piper and Piper X arms as diagnostic material, and 1 of them was a full success. The authors write that Astra interpreted tasks more reliably than it executed them safely, and they call for independent safety checks and stop mechanisms instead of trust in the model's judgment.
The incidents are described in text; the paper has no video or logs of them.
Written by the Boards & Bots curator
“Unsafe actions halted our real-robot trials.”