Asimov 1 is an open-source humanoid anyone can extend
Open-source humanoids have mostly meant open-source marketing with a closed product underneath. Asimov 1, a 1.2 metre, 35 kilogram biped from Menlo Research, is the exception. The mechanical CAD, electrical design, simulation models, firmware, reinforcement-learning training code and the Python SDK are all public. What that has started to produce is outside work. On Sunday morning a developer with no connection to the company published a runtime safety experiment he had built on the project's own simulation model, complete with pre-registered results, and filed it as an issue. An open pull request against the LeRobot robot learning framework would add Asimov 1 locomotion through the Menlo SDK's UDP transport, exposing 25 measured joint positions along with the robot's IMU and camera feeds. The official repository lists 25 actuated degrees of freedom, a 7075 aluminium and MJF-printed nylon structure, a Raspberry Pi 5 handling media and networking, and a Radxa CM5 running motion control, with 1,532 stars and 249 forks at the time of writing.
The project's own team has been widening the published surface at the same time. The training framework that produces the walking policy is a standalone Isaac Lab extension running proximal policy optimisation with adversarial motion priors, and on 7 November it gained a new physics option that lets the locomotion tasks run on MuJoCo Warp through Isaac Lab's Newton backend while keeping PhysX as the default. Three days earlier a merged pull request added printable accessory parts including TPU shoe soles and a dummy hand set, the kind of small addition that only makes sense once the rest of the design is already public. What the Sunday issue shows is what the published surface now makes possible. A stranger could load the published MuJoCo model, read the published joint gains, and test a safety layer of his own design against it.
The reason the project exists is more prosaic than any of that, and it explains the design choices. Menlo's own handbook describes a team that bought a commercial humanoid to run their own software on, deployed a walking policy to it for the first time, and watched the machine lose control and drive its leg into the floor, shattering a small metal part in the ankle. The company then stood still for two months waiting for one replacement component to arrive. From that experience came the rule that a robot you cannot repair yourself will hold you back, and eventually the decision to build their own: legs first, then a torso, then a whole body, with the designs published as they went.
What the open-sourcing buys in practice is concrete and slightly unusual. The hardware sits under CERN's Open Hardware Licence version 2 in its strongly reciprocal variant, and the April announcement put the onboard software under GPL-2.0 while the training code carries a BSD-3-Clause licence, so a builder can commercialise a derivative of the training stack without inheriting the copyleft terms of the firmware. The documentation shows two compute stacks rather than one. The first is a head unit holding the Pi 5 with a media board that handles the buzzer and brownout protection. The second is a robot processing unit carrying the Radxa carrier board, the CM5 and the power distribution board, with the carrier plus module forming the motion control board. The joints talk over six CAN buses, five at 1 Mbps and one at 500 kbps, and the repository's specification table lists a 5 kilogram squat load and 18 kilogram peak lateral raises on each arm.
The walking policy is where the open-sourcing matters most, because a walking controller is normally the one thing a humanoid vendor keeps to itself. The released policy ships as an ONNX file under a permissive licence, and according to CircuitDigest it runs at 50 Hz, reading 78 values every 20 milliseconds and emitting target positions for 23 joints, with the neck held locked during locomotion. The training repository exposes the reward definitions and the joint-specific actuator settings, so the penalties for foot slipping, abrupt action changes and self-collision are all readable, along with the torque limits, stiffness, damping, friction and actuator delay for each joint. The simulator also varies foot friction, starting joint positions and actuator gains, and injects disturbances and noisy observations to keep the learned gait from depending on one idealised world. Humanoids Daily walks through the published configuration. Training runs on a single GPU with 4,096 parallel environments, and the menlo-sdk package puts walking, joint control, state, camera and audio behind a Python interface whose every command passes through the same arbiter and safety layer on the robot as the gamepad.
The safety experiment in that Sunday issue is the clearest evidence that the openness is real rather than promotional. The author, who describes himself as working on a runtime guardian that sits between a planner and the motors, loaded the published MuJoCo model unmodified, fixed the pelvis on a stand, and added position actuators using the project's own PD gains. He ran eight scenarios across twenty random seeds each with the guardian on and off over identical command streams, with the metrics and the video rule fixed in advance. The results were blunt: in the scenario where a planner reaches into a seated mannequin's torso, contact went from 20 runs out of 20 to none, and peak simulated contact force went from 157 newtons to zero. For the fast swing toward a head it was 20 out of 20 down to none and 810 newtons down to zero. Fuzzing went from 16 out of 20 to none. The check itself cost a median of 0.46 milliseconds per control step with a 99th percentile of 2.8 milliseconds, running on single-threaded Python on a laptop rather than the CM5 that would actually sit on the robot. The caveats he lists sit alongside those numbers: the pelvis was fixed so balance is untested, the hands are not modelled, and the guardian has no table keep-out, which is why the arm still hit the table in 18 of 20 fuzzing episodes even with it on.
The questions he raised in that issue are still open in the repository, namely whether joint velocity limits are published anywhere, since the MJCF carries position and torque limits but no velocity ones, and where in the stack a command filter would sit so that a safety layer can be added without cutting torque on a block mid-motion. Meanwhile the hardware side is already shipping: the first batch of DIY kits went out in late August at 20,000 dollars delivered, a price the team reached after finding that standalone actuators of this class cost upwards of 30,000 dollars when bought one at a time at low volume, and then negotiating a group buy that pulled prices near factory cost. An integration into someone else's robot learning framework is the kind of thing that used to need a vendor's cooperation, and the LeRobot pull request needs none. Whether strangers can use it without asking permission is the practical test of an open platform, and on that measure Asimov 1 is already past the interesting part. In the same issue the author also asked which licence covers the simulation directory, and offered to contribute a small out-of-tree adapter if the answer is workable.
Featured blog posts
A Transformer model called EarlyDetect, trained on SDO/HMI helioseismic and magnetic observations by an NJIT-led team including Princeton and NASA Ames, forecasts the emergence of solar active regions on the Sun an average of 9.24 hours before they become visible, outperforming LSTM baselines and revealing that temporal smoothing filters hid the faintest precursor signals (Tirona et al., JGR Machine Learning and Computation, 14 August 2026) Posted on May 24, 2026
Posted on May 4, 2026
Today we are joined byYasunori Yamazaki, Chief Business Officer at Axelspace. Axelspace are pioneers of microsatellite technology advancing the frontiers of space business, reimagining traditional ways of using space, and creating a society where everyone on our planet can make space part of their life. Latest blog posts
"Menlo Research's Asimov 1 is a 1.2 m, 35 kg, 25-joint humanoid whose mechanical CAD, electrical design, MuJoCo and URDF simulation models, firmware, reinforcement-learning training code and Python SDK are all published under CERN-OHL-S-2.0, GPL-2.0 and BSD-3-Clause, and by early November 2026 outside developers were submitting work against it, including a runtime safety layer experiment and a pull request adding the robot to the LeRobot framework"'
Liu, Saper and colleagues report in Science that Teghaza 001, a 798.5 gram Martian meteorite recovered in Mali in 2022, is a slowly cooled igneous rock formed at least 4.1 billion years ago that samples an unsampled part of Mars' early crust and preserves a chemical record of water rich fluids and of hydrogen lost to space
A team led by Anna Karpova and James Taylor used the Green Bank Telescope to find PSR J2053+6905, a 1.68 millisecond binary pulsar and the third fastest redback known, whose optical light curve brightened on one side between 2023 and 2025 and whose mass of 2.12 solar masses constrains the equation of state of ultradense matter.
NASA's announcement gives a takeoff weight of 10,000 pounds, a Mach 0.75 cruise and room for 2,000 pounds of swappable experiment, but leaves out the engine count, the wing geometry and the contractor. What the airframe is actually for is finding out how much control authority a crewed airplane can be given before the ordinary surfaces stop working.
United Launch Alliance's Vulcan Centaur suffered an anomaly 10 minutes into its launch on 15 November 2026 from Space Launch Complex 41, destroying 40 Amazon Leo satellites on the inaugural Leo Vulcan 1 mission and marking ULA's first total mission failure since the Boeing and Lockheed Martin partnership was formed in 2006, with the failure occurring during the hydrogen-burning Centaur upper stage's planned 12-minute 15-second burn after the solid boosters worked normally and the Blue Origin BE-4 first stage appeared to burn longer than its pre-launch timeline expected
"Protopapa et al. 2026 in The Astrophysical Journal Letters report that JWST/NIRSpec spectra of interstellar comet 3I/ATLAS taken in August 2025, December 2025 and April 2026 show the coma's water ice locked into sub-micron aggregates loaded with rocky material, which suppresses the 1.5 and 2.0 micron water-ice overtones while keeping the 3 micron band deep, placing 3I/ATLAS at the dusty end of a continuum of ice grain states shared with trans-Neptunian objects"'
SpaceX caught the Starship upper stage Ship 42 between the arms of Orbital Launch Tower 2 at Starbase at 21:20:03 UTC on 14 November 2026, the first catch of an orbital rocket upper stage returning from space, after 22 hours 32 minutes and 14 orbits in which the ship deployed 26 Starlink V3 satellites and was photographed in orbit by three of them, while Super Heavy Booster 22 lost its own catch when a low fuel mass alarm ended the boostback burn early
Lithium fluoride crystals flown inside two human-shaped phantoms on the 2022 Artemis I mission show that radiation dose near the Moon exceeds low Earth orbit by less than expected, because the vanishing Van Allen proton contribution is offset by galactic cosmic rays that pass through a body. The dose also comes out nearly uniform through a torso, which changes how lunar shelters have to be built.
SpaceX's fifteenth Starship flight lifted off from Starbase at 22:47 UTC on 13 November 2026 as the first attempt to catch both stages of an orbital rocket with the launch tower's mechanical arms, but Super Heavy Booster 22 was directed to an offshore splashdown in the Gulf of Mexico about six and a half minutes after liftoff instead of being caught, while Ship 42 reached a 275 km orbit, deployed all 26 Starlink V3 satellites, and began a roughly 21 hour coast before a first ever attempt at catching a Starship upper stage at the tower, scheduled for about 21:20 UTC on 14 November
A coordinated Hubble and Webb survey found 27 previously unknown trans-Neptunian objects, the faintest about 10 km across, at a depth deeper than any previous Solar System survey, and found that their surfaces and size distribution preserve the composition and formation history they had at the birth of the solar system
Subscribe to blog posts using RSS











