Nvidia and Pollen Robotics train $399 Microduck robot entirely in simulation
Nvidia and Pollen Robotics showed that a locomotion policy for the $399 Microduck robot can be trained purely in Isaac Lab 3.0 and transferred to real hardware with no tuning. Training ran 4,096 parallel environments on one GPU at up to 1.25 million steps per second.
- Microduck: 25 cm tall, ~780 g, 15 hobby servos and an RK3566 with a 1 TOPS NPU
- $399 pre-orders passed 10,000 units, generating about $4 million in revenue
- Isaac Lab 3.0 models servo backlash, friction and 10–30 ms command latency
- Training used 4,096 environments on one GPU and scaled across 16 GPUs
Read next
Hardware