What happened
CircuitDigest reported on October 28, 2026 that Menlo Research brought its Asimov 1 humanoid to the SWITCH event in Singapore on day 366 of the project. The 1.2-metre, roughly 35-kilogram robot reportedly walked among visitors using a newly trained locomotion policy. The article says the project’s GitHub repository had passed 1,300 stars.
Asimov 1 is presented as an open hardware and software project rather than a closed commercial humanoid. Menlo has published mechanical CAD, electrical schematics and PCB files, simulation models, firmware, and the walking policy. The hardware licence is CERN-OHL-S-2.0 and the software licence is GPL-2.0. A DIY kit is listed, while a full SDK and emulator remain under development.
The robot has 25 powered joints: six in each leg, five in each arm, one waist joint, and two neck joints. Its structure combines 7075 aluminium with MJF-printed PA12 nylon, and it uses a 13S4P lithium-ion battery pack. The compute design is split: a Radxa CM5 runs motion control, while a Raspberry Pi 5 handles media and networking. Six CAN buses connect the joints, with five operating at 1 Mbps and one at 500 kbps. Reported sensing includes an IMU, a 2 MP camera, a stereo microphone array, and a 10 W speaker.
The walking controller is a neural-network policy trained in NVIDIA Isaac Lab with PPO reinforcement learning and adversarial motion priors. At runtime, the policy operates at 50 Hz. Every 20 milliseconds it receives 78 values, including body angular velocity, gravity direction, commanded speed, joint positions and velocities, and the prior action; it produces target positions for 23 joints. The neck is locked during walking. Menlo has released checkpoints as ONNX and TorchScript files on Hugging Face alongside training code. According to the report, policies trained with the project’s MuJoCo and URDF models can be deployed to the physical robot without additional tuning. The CM5 firmware image already includes the walking policy, enabling walking from velocity commands, while a direct joint-control mode is available for teleoperation.
What the archive adds
This is a new fact on a known story: the earlier Asimov-1 v1 Field Note established Menlo Research’s project as distinct from RoboParty’s RP1. The reported public walk adds an operational demonstration of a newly trained policy, rather than another description of the platform’s initial release.
Trend signal
No supplied pattern directly fits this hardware-and-locomotion update.
Connections to other events
The development sits alongside recent open humanoid-platform coverage. PhoneBot explored reducing component requirements by reusing a smartphone for sensing and high-level computing. RoboParty RP1 emphasized inspectability across hardware, firmware, deployment, and learning. Asimov 1 similarly exposes multiple layers, while its reported CM5-plus-Pi split makes the control boundary explicit.
Why it matters
A walking video alone does not make a humanoid reproducible. The more consequential part of this release is the claimed chain from simulation assets and reinforcement-learning code through portable checkpoints, firmware, and a physical control stack. That chain gives university labs, robotics clubs, and small engineering teams artifacts they can inspect when debugging sim-to-real behavior instead of treating locomotion as a vendor capability.
The compute split is also a practical design choice to study. Motion control runs separately from media and networking, reducing the temptation to put timing-sensitive actuation beside less predictable application workloads. Builders should not infer that the architecture is universally sufficient, but it is a concrete reference point for separating real-time control from higher-level interaction functions.
What happens next
- Menlo releases a fuller SDK or emulator for Asimov 1. (6 months, 60%). Watch for: a public repository, documentation update, or versioned release.
- The project publishes additional policies or evidence about robustness beyond the event demonstration. (6 months, 55%). Watch for: new checkpoints, evaluation conditions, or repeatable deployment results.
Limits and uncertainty
The supplied report describes a public demonstration but does not provide a standardized locomotion benchmark, fall rate, recovery behavior, duration of operation, battery life, or comparison against other humanoids. It does not establish how reliably the policy performs across surfaces, payloads, disturbances, or different physical builds. The statement that simulation-trained policies transfer without extra tuning is attributed to Menlo through the report and is not independently evaluated here. A walking policy and direct teleoperation also do not demonstrate autonomous manipulation, perception-led navigation, or safe operation in unconstrained public environments.
Skill to build now
Treat a robot policy as a deployable artifact with an explicit interface. This week, document the observation vector, action vector, rate, actuator mapping, simulator model version, and firmware version for one existing control loop. Export one checkpoint into a portable runtime format such as ONNX where appropriate, then verify that an identical command path can be exercised in simulation and on hardware. That discipline makes later claims about transfer testable rather than anecdotal.
Reader takeaway
Asimov 1’s reported public walk matters because its creators have exposed more than a finished robot: the mechanical, electrical, simulation, firmware, and policy layers are intended to be available for inspection. The evidence is still a single reported demonstration, but the release offers a concrete open stack for builders studying humanoid locomotion.
Open question
Will independent builders be able to reproduce the reported sim-to-real walking result across their own Asimov 1 hardware, without substantial unstated calibration or tuning?