IN Brief:
- LG’s next bipedal humanoid will use NVIDIA Jetson Thor for onboard compute alongside Isaac GR00T and Halos for Robotics.
- LG plans to combine actuators, sensors, and batteries developed across its own group companies.
- CLOiD robots will be validated on an LG washing-machine production line in Tennessee before wider deployment.
LG is developing a bipedal humanoid around NVIDIA’s Jetson Thor embedded computing platform and Isaac GR00T robotics model, while using factory deployments to gather the operating data needed to refine later hardware and software.
The robot is scheduled for a public unveiling in the first quarter of 2027 and will use Jetson Thor for onboard compute, reasoning, and control. LG is also incorporating NVIDIA Isaac GR00T and Halos for Robotics, while drawing on actuators, sensors, and batteries developed across LG Electronics, LG Innotek, and LG Energy Solution.
The combination places a substantial electronics integration problem behind the mechanical form of the humanoid. A mobile robot has to run perception, localisation, AI inference, motion planning, joint control, communications, and safety functions within one power and thermal envelope, while maintaining sufficiently low latency to keep physical movements stable.
Jetson Thor is intended to keep more of that processing on the machine rather than depending on constant communication with an external server. Local compute reduces network latency in control loops and keeps perception and movement decisions closer to the sensors and actuators producing and consuming the data.
The resulting processor workload can vary sharply. Camera and other sensor streams have to be interpreted continuously, while motor-control systems operate at different frequencies and with more deterministic timing requirements. Memory capacity, accelerator utilisation, I/O bandwidth, and thermal limits consequently have to be considered as parts of the same architecture rather than separate subsystem choices.
LG also plans to integrate hardware supplied from within the wider group. Bringing actuators, sensing, battery technology, and compute into one programme may give engineers more control over mechanical packaging, power distribution, diagnostics, communications, and control-loop behaviour than would be possible with an entirely off-the-shelf collection of subsystems.
Battery design is particularly difficult for a mobile humanoid because the electrical load includes both high-current electromechanical movement and comparatively continuous computing. Peak actuator demand can change rapidly with posture and movement, while the AI processor, sensing, communications, and control electronics create an always-on base load.
That leaves engineers balancing cell capacity against mass, current delivery, thermal behaviour, operating time, and charging strategy. Adding battery capacity extends runtime but also increases weight, which can raise the energy required to move the robot and place greater loads on its actuators.
LG and NVIDIA are coupling the hardware programme with a robot data environment based on LG CNS PhysicalWorks. The planned infrastructure will support continuous collection of physical-world data, synthetic-data generation, training, and verification rather than treating AI development as a one-off exercise completed before the machine reaches a site.
Real production data will come partly from LG’s CLOiD wheel-based robot. The company plans to deploy CLOiD on an LG Electronics washing-machine manufacturing line in Tennessee during 2026, using the site to validate performance before considering broader deployment.
A production line provides a less forgiving test environment than a staged robotics demonstration. Lighting changes, people move through shared space, equipment creates physical obstructions, production conditions vary, and the robot has to operate without becoming an unpredictable interruption to an established manufacturing process.
Those conditions can expose weaknesses that affect the electronics architecture directly. Repeated sensor occlusion can lead to changes in sensor positioning or modality, unexpected actuator duty cycles can alter motor-drive and thermal requirements, and simultaneous perception and control workloads can reveal memory or processor bottlenecks that are less obvious during short laboratory tests.
The collaboration is intended to feed that operating data back into LG’s own Robot Foundation Model as well as its use of Isaac GR00T. The compute platform therefore needs enough flexibility to accommodate changing AI workloads rather than being designed around one fixed network and frozen for the life of the robot.
LG and NVIDIA have established a joint technical and commercial task force covering R&D, site validation, and eventual commercialisation. That becomes important as the programme moves beyond prototypes, because repeated production introduces conventional engineering requirements around component sourcing, assembly tolerances, calibration, production test, firmware control, repair, and configuration management.
A technically capable robot that requires extensive expert adjustment after every assembly is still a poor manufacturing design. The Tennessee validation programme should provide LG with evidence about which parts of the architecture can tolerate real operating variation and which remain dependent on laboratory conditions.
The planned 2027 humanoid unveiling will provide the visible milestone, but the production-line work is likely to be the more useful engineering test. If the same compute, sensing, actuation, battery, and software architecture can be reproduced and maintained across several operating environments, LG will have moved considerably further towards an industrial robot platform rather than another carefully prepared humanoid demonstration.
