Alpamayo 2 Edge

Methodology

• Latency: Jetson AGX Orin 64 GB (MAXN) running this build and its server on the 16 drives below, 100 decisions each, back to back; the first 10 of each drive are left out (1440 decisions). Latency, model: camera frames to trajectory, as the server reports it (median 88.2 ms, p90 89.1 ms; per-drive medians 87.9 to 89.0 ms). Throughput, model: 1000 / (the mean model time plus the mean work the server does between decisions). Latency, end to end: request sent to reply received on the device, request decoding included (median 92.5 ms). Throughput, end to end: decisions answered per second, back to back, on the same 16 drives. Before each drive, 21 untimed decisions warm the server up; they are not shown or counted. The reasoning in words is refreshed after the reply and used by the next decisions, so it does not add to the latency. The action expert takes 2 steps instead of 10. INT8, except the last two language-model layers, which stay FP16 to keep the reasoning.
• Accuracy: minADE6 at 5 s: the closest of 6 sampled trajectories, average error over 5 s, on 189 held-out PhysicalAI-AV validation clips, 2 seeds. Embedl on the device, NVIDIA BF16 on an H200, same clips. Embedl's error is 20.6 % higher (95 % CI +8.4 to +33.2 %): this build trades accuracy for speed. It reads 7 of the 22 camera frames at fewer pixels and runs a 2-step action expert; on 2,000 other validation clips (BF16, PyTorch) its error is 24.5 % higher than NVIDIA's full-input model's. Measured with the reasoning decoded for each decision.
• Inputs and fine-tune: To fit 10 decisions per second, this build reads 7 of the 22 camera frames (the newest frame of each camera, and the front wide camera 1 s earlier) at 24 to 60 % of NVIDIA's pixels. Embedl fine-tuned NVIDIA's released weights for these inputs and the 2-step action expert, learning from NVIDIA's own outputs on PhysicalAI-AV training clips only. It is less accurate than NVIDIA's full-input model (see Accuracy); the all-frames Embedl build keeps NVIDIA's inputs and accuracy at about 3 decisions per second.
• Power and memory: Power: the board's power rails (tegrastats, every 200 ms) while it decides back to back. Energy per decision: that mean power times the median model time. Idle: the board at rest. Peak memory: the board's RAM in use at the peak, minus before the model started; the GPU shares this memory.
• Size: The runtime container image and the model files on disk.
• TensorRT Edge-LLM FP16: NVIDIA's released weights in FP16 with all 22 camera frames, on the same runtime and device, without Embedl's quantization or fine-tune. Shown for speed.
• NVIDIA PyTorch BF16: NVIDIA's released code (NVlabs/alpamayo2-edge) and BF16 weights (nvidia/Alpamayo2-Edge) with all 22 camera frames, on an H200. Shown for accuracy.