A $2 ESP32-C61 and an 8 GB Jetson | What changed in AI robotics

What actually changed in AI robotics over the last two days? On humanoids, nothing you can check against a primary source. What did move is the layer underneath, the one you touch: an Espressif Wi-Fi 6 module now selling for around $2, that same maker open sourcing its entire cloud stack, and two papers that lower the machine you need to train and to deploy. Same thread running through all of it, the entry ticket gets cheaper.

The ESP32-C61-MINI-1 now sells for about $2

Careful with the word “new” here. The C61 silicon has been around for a while. What changed is that CNX Software documented on 4 August a 5-pack of MINI-1 modules at $10.70, so roughly $2 per module. Module, not bare chip: it ships with the antenna and the certification, which saves you the RF work.

What Espressif’s datasheet states: single-core RISC-V up to 160 MHz, 320 KB SRAM, up to 8 MB Quad SPI flash, up to 8 MB optional PSRAM, Wi-Fi 802.11ax on 2.4 GHz at 1T1R, Bluetooth LE 5.3 and up to 23 available GPIOs. The MINI-1 measures 13.2 x 16.6 x 2.4 mm with a PCB antenna, the MINI-1U swaps it for an external antenna connector. Operating range is -40 to 85 °C.

What is it for inside a robot? The telemetry node, the remote sensor, the handheld controller, the leg that reports back to the main board. One core at 160 MHz and 23 GPIOs is not the brain of anything complicated, and keep in mind that Wi-Fi 6 here is 2.4 GHz only. The point is cost per node: at $2 you can drop three into a project without thinking hard about it.

And the usual warning. That module gives you milliamps on a pin. A servo asks for hundreds. Separate supply and a driver in between, or you fry it.

Espressif open sources all of RainMaker Neo, Apache 2.0

This is the 5 August item. Espressif posted on its developer blog that ESP RainMaker Neo goes open source under Apache 2.0, and not only the firmware. Per the announcement, what opens up is the full device to cloud to app stack: the cloud backend, the device firmware, the mobile apps and their SDKs, the dashboards, and the Alexa and Google Home voice assistant integrations.

The small print is in the announcement too. Neo is not Classic with a new name: it is a leaner implementation that leans more directly on native AWS IoT services, and Espressif says it is not a migration path for existing Classic deployments. So it is for starting fresh, not for moving house.

What does that give you while building robots? The connectivity and remote control layer stops being somebody else’s closed box. If you have an ESP32 robot you want to drive from your phone, or a node reporting telemetry, you can now read the backend, change it and host it yourself. Keep in mind that leaning on AWS means the cloud bill is yours: open source does not mean free hosting.

An 8 GB Jetson can in fact run a bimanual policy

This is the most useful find of the two days. A team published a study on 4 August putting a full bimanual SO-101 system inside an NVIDIA Jetson Orin Nano Super with 8 GB. The RTX 3070 is used only for offline training. In other words, the expensive part stays on your desk and what rides on the robot is the small board.

The numbers the authors report, and this is a preprint with no peer review: building the three-camera capture on NVMM buffers without redundant copies drops peak single-core CPU from 98% to 77% and worst-case latency from 117.31 ms to 101.52 ms. Converting ACT to TensorRT takes mean latency from 114.02 ms down to 17.93 ms in FP16 and 12.65 ms in INT8, with task success holding. They also report something that rarely gets written down: ACT reached 19 of 20 trials while Diffusion Policy never converged, 0 of 10, on the same data.

Translated to your bench: on a small board the bottleneck is often not the model, it is how you move camera frames around. And quantizing is not a last-minute trick, it is what makes the deployment viable at all. I found no public repo for the paper, so for now it is an approach worth copying, not a project you clone this afternoon.

Training an action model on a single 24 GB GPU

The second paper from the same day, LiLa-WAM, attacks the other end: training. They propose a world-action model that works in a compact latent space and trains end to end on a single 24 GB GPU. They report 90.48% success across 50 RoboTwin tasks.

Time to slow down. That 90.48% is the authors’ own figure and it is simulation, not your living room. RoboTwin is a benchmark, and a benchmark measures what it measures. The interesting part is not the number, it is the machine budget: a second-hand 3090 or 4090 fits in that range, and that moves the line on who gets to train something like this.

What I would do today

If you are building: note the C61-MINI-1 for your dumb nodes and leave the brain where it is. If remote control is what is blocking you, look at RainMaker Neo before you write your own server. And if you are working on vision or manipulation and have a small Jetson gathering dust, read the 8 GB paper and start with the boring part, zero-copy capture and the TensorRT conversion. Sounds small, I know. That is exactly where the headroom was.

And no, still no new humanoid. Across these two days no locomotion or manipulation announcement from the big makers could be checked against a primary source. When there is one, we will run it with the link right next to it.