Run #LLMs# on @NVIDIARobotics #Jetson# without x86 ONNX export. 🚀 #TensorRT-Model-Connect# on reComputer Classic J5012 (#AGXOrin# 64GB, #JP7#.2). @huggingface checkpoint → trtmc build → .bundle on device trtmc run for text gen Tested: #Qwen3-4B# FP16 (~41GB container peak, 64GB recommended) No ONNX hop. Edge LLM/agent ready. Would you use on-device TRTMC for prototyping, or keep ONNX export for production? 👇 Full guide: Get your own one: #NVIDIA# #Jetson# #TensorRT# #TRTMC# #AGXOrin# #EdgeAI# #LLM# #PhysicalAI# #SeeedStudio#

Today I unboxed a Jetson AGX Orin 64GB. This is the on robot box. Cameras, CAN, capture, Isaac ROS. The nervous system for whatever I bolt it to. The rest of the lab: 2x DGX Spark for GR00T train and serve (Isaac Sim runs here too) RTX 4080 / 64GB DDR5 for Sim, demos, and Cosmos Physical AI, day 1. @NVIDIARobotics

Robotics deep dive: HUMANOID ROBOT EMBEDDED COMPUTE Every robot company that says anything about its onboard compute quotes a TOPS number. Almost none of them are quoting the same thing. NVIDIA's Jetson Thor markets 2,070. The footnote on NVIDIA's own page: sparse FP4. The previous generation, AGX Orin, markets 275, under a column header reading sparse INT8. Divide one by the other and you get the headline claim for the generation, up to 7.5x more AI compute. Put both on the same footing and it is 3.76x. Dense FP8 against dense INT8, 517 against 137.5. The other half of the speed-up is the number format. NVIDIA states 3.5x better energy efficiency, which reconciles as 2,070 over 130 watts against 275 over 60. At matched precision it is 1.74x. Both headline claims collapse by exactly 2x, by the same mechanism. Sparsity is real, FP4 is real, and NVIDIA prints both in its spec table. Thor is a large generational step. It is roughly half the step the marketing arithmetic describes, and the difference lives in a footnote most people never resolve. Unitree's page for the H2 PLUS states its chip correctly: FP4 2,070 TFLOPS. Search-indexed writeups of the same robot render it as a 2,070 TOPS chip. Floating point silently became integer and the precision qualifier vanished. The spec survived the vendor and died in the coverage. Qualcomm's Dragonwing IQ10 markets up to 700 TOPS and states no precision. Horizon's Journey 6P markets 560 and states none either. Either could be INT8 or INT4, a factor of two, with no document that settles it. Neither publishes a power figure, so neither can be placed on a per-watt axis at all. Horizon does publish one thing nobody else does. The footnote under its TOPS figure reads TPP under 4800. Total Processing Performance is the threshold metric in US export control. That number is bounded by a regulation rather than by the silicon. Figure, 1X, Agility and Apptronik publish no compute specification at all. The only two humanoid vendors that name their chip in a public spec table are Chinese.

NEW UNITREE COMPETITOR JUST DROPPED! Booster Robotics @boosterobotics is a Beijing humanoid-robot startup founded in June 2023 by CEO Cheng Hao, with a core team from Tsinghua's Robot Control Lab and its Hephaestus RoboCup soccer team. It sells small, walk-and-run-optimized bipedal humanoids as a research, education and competition platform rather than for industrial labor. Its product line consists of: - the K1 (about 95 cm, 19.5 kg, 22 DoF, NVIDIA Jetson Orin NX, ROS2 and Isaac Sim support), starts at $4,999. - The flagship T1 (about 118 cm, 30 kg, 23 DoF, Jetson AGX Orin, up to 18 km/h), runs roughly $22,000 to $34,000. - The Professional T2 (July 2026), that uses NVIDIA's Thor compute. It ships a developer IDE, Booster Studio, and its robots won both humanoid leagues at RoboCup 2025. Booster Robotics targets research, education and competition, has shipped more than 700 units to over 200 customers (more than half internationally), and reached positive monthly operating cash flow in December 2025, while the marquee humanoid companies (Figure, Tesla Optimus, even Unitree pre-IPO) burn cash chasing a factory-labor future. Booster already monetizes the developer ecosystem (cheap robots plus the Booster Studio IDE plus ROS2 and Isaac Sim) instead of promising labor replacement later. -> It is the anti-Figure play. Both 2025 RoboCup humanoid league champions ran on Booster hardware! Tsinghua's Hephaestus won AdultSize on the T1 and Germany's Boosted HTWK won KidSize on the K1 Winning robot soccer is an unfakeable test of dynamic bipedal control (unstructured, adversarial, contact-rich, you cannot cherry-pick a match), and sweeping both size classes, including a German team choosing a Chinese robot, means the serious humanoid-research community standardizes on Booster. Pricing is not random. The K1 at $4,999 puts a walking humanoid in a classroom and undercuts even Unitree's G1 (about $16,000); the T1 at roughly $22,000 to $34,000 is the RoboCup-grade research machine; the T2 (July 2026, NVIDIA Thor, a claimed ~2,070 TFLOPS) is the pro edition. Sub-$5,000 for a real biped is a bet on the next generation: whoever learns robotics on your platform builds for it. Dexterity is deliberately not their game. The T1 is optimized for walking and running (18 km/h, omnidirectional) and won a soccer league, and Booster describes its IP as a motion-algorithm platform that decouples control across configurations. The compute is NVIDIA (Jetson Orin, Thor), the stack is ROS2 and Isaac Sim, and Booster Studio (June 2026) adds agent development, "vibe coding," simulation and one-click deployment to the physical robot. More than half of units ship internationally, so this is China exporting a developer platform, not just hardware, and riding the global open-robotics toolchain rather than a closed domestic one, the inverse of the usual worry about vertically closed Chinese stacks. Backers include the Beijing Robotics Industry Development Investment Fund, the Beijing Artificial Intelligence Industry Investment Fund and Shenzhen Capital Group, that is, municipal state money seeding a humanoid-developer champion.