Preparing the published article.
Astribot: S1 Manipulation Robots, Research Toolchains and Robot Learning
Shenzhen robotics developer combining S1 hardware with teleoperation and embodied AI research. Review model-specific specifications, development access, Lumo-2 and SmoothRL evaluation boundaries.
Company and development approach
Astribot was founded in Shenzhen in December 2022. The company develops humanoid robotics and describes its approach as Design for AI, combining hardware and software for manipulation and robot learning. Its S1 platform, development tools and research projects belong to one supplier; the scope of this profile is the documented S1 research platform and associated learning work.
S1 hardware and specification scope
The official S1 research specification lists seven degrees of freedom per arm, a five-kilogram payload per arm at horizontal reach, and a height of 170 centimetres. These figures describe that published configuration. End-effector speed is an arm-motion measure, not a walking-speed rating, while arm span and repeatability should not be interpreted as a guaranteed task workspace or finished-process accuracy.
Development tools and teleoperation
Astribot presents APIs, development guidance, a visual interface, simulation support and a VR teleoperation approach for data collection. Before choosing a platform, identify the delivered software versions, control interfaces, simulator assets and documentation. Confirm which functions are available with the purchased configuration and which require additional access, integration or an agreement with the supplier.
Lumo-2 robot-learning research
The Lumo-2 project describes a model that combines latent world dynamics with action generation and progressive alignment across modalities. Its research explores robot demonstrations, human video and other data sources, with reported evaluations of manipulation and transfer. Those research results should be read with their task definitions and evaluation settings; they do not establish reliable performance on every workplace task.
SmoothRL and human intervention
SmoothRL addresses online reinforcement learning during asynchronous robot operation. Astribot reports experiments on S1 involving object tossing, pen capping and box opening, and describes both VR takeover and joystick-based corrections. Human intervention is part of the presented learning setup. Reported improvement within these experiments should remain separate from a claim that a deployed system needs no supervision.
Pilot acceptance and data handling
Define a pilot around representative objects, lighting, reach limits and the actual work sequence. Record successful completions, failures, operator assistance, recovery time and changes to the environment. Agree how demonstrations and operational data are stored, exported and reused, including responsibility for access controls. Compare results over repeated trials before expanding the task or operating area.
Delivery and support questions
Request the exact hardware and software bill of materials, commissioning scope, training plan and service arrangements. Confirm replacement-part availability, update support, permitted operating conditions and the process for recovering from a failed experiment. Include computing, teleoperation equipment and recurring software costs in the proposal. A product demonstration should lead to a documented acceptance plan for the intended facility.