This paper deploys a bimanual SO-101 manipulation system entirely on an 8 GB NVIDIA Jetson Orin Nano Super, using an RTX 3070 only for offline training. A three-camera GStreamer pipeline with NVMM buffers reduces peak single-core CPU utilization from 98.0% to 77.0% and worst-case latency from 117.31 ms to 101.52 ms. ACT succeeds in 19/20 beanbag pick-and-place trials, while Diffusion Policy achieves 0/10 under the reported training budgets. TensorRT reduces ACT inference latency from 114.02 ms to 17.93 ms in FP16 and 12.65 ms in INT8, with success rates of 18/20 and 19/20, respectively.
No heat snapshots are available in the last 24 hours.