By Sarah Xiao 
What Is the Current State of Humanoid Robot Planning Capabilities?
From Boston Dynamics Atlas to Tesla Optimus, from UBTECH Robotics Walker to Xiaomi CyberOne, humanoid robots have suddenly become one of the hottest sectors in the technology world over the past two years.
Capital has poured in, major corporations are racing to establish their positions, and the entire supply chain—from upstream components to downstream applications—is buzzing with activity.
But is this truly a technological revolution, or merely another tech bubble waiting to burst?
Breakthrough or Buzzword? The True Level of Current Progress
Phased Advances in Hardware Technology
From a technical architecture perspective, humanoid robots have indeed achieved substantial, tangible progress.
By 2025, Tesla Optimus Gen 3 had reportedly entered mass production, with each robot featuring 28 degrees of freedom and using Tesla’s in-house actuators and torque sensors. This marked a key milestone in the transition of humanoid robots from laboratory prototypes toward industrialized production.
After years of iteration, Boston Dynamics Atlas has demonstrated the ability to perform highly complex movements such as backflips and parkour, showcasing exceptional balance control and whole-body coordination.
Meanwhile, Figure AI Figure 02 demonstrated real operational capability in a BMW factory in late 2024, independently performing parts handling and assembly tasks.
However, it is crucial to recognize the significant gap between demonstration environments and real-world deployment. Most demo videos are recorded in carefully staged settings—flat floors, adequate lighting, and highly standardized tasks. In real industrial environments, conditions are far more complex: floors may be oily, lighting insufficient, and tasks often require on-the-spot judgment. These factors place much higher demands on a robot’s perception, decision-making, and execution capabilities.
Software Algorithms: Bottlenecks and Progress
The level of intelligence in humanoid robots largely depends on their software systems. The mainstream technical approach today combines large language models (LLMs) with robotic control systems, enabling robots to understand natural language instructions and translate them into physical actions.
Tesla’s approach draws on its experience in autonomous driving, adopting an end-to-end neural network training paradigm. Visual input from cameras is fed directly into neural networks, which output joint control commands. This approach offers strong generalization capability, but requires massive training datasets and suffers from poor interpretability.
Figure AI, by contrast, adopts a modular architecture, separating perception, planning, and control into layered subsystems. This approach aligns more closely with classical robotics theory and offers better interpretability, but tends to be less adaptable in highly complex or unstructured environments.
Despite ongoing advances, current AI technology still struggles to fully solve the robot’s “common sense problem.” Humans can easily judge whether an object is fragile, whether a surface is slippery, or whether a structure is stable—judgments that robots can only acquire through extensive training. More importantly, humanoid robots must make real-time decisions in dynamic environments, placing extreme demands on computing power and system latency.
Energy and Endurance: A Hard Constraint
The energy system of humanoid robots is a critical yet often underestimated bottleneck. Most current humanoid robots rely on lithium-ion batteries, with typical operating times ranging from 2 to 4 hours. Under an 8-hour work shift, this implies multiple charging cycles, significantly reducing operational efficiency.
Tesla Optimus Gen 3 is equipped with a 2.3 kWh battery pack, offering a theoretical runtime of approximately 4 hours. In real operating conditions, however—due to frequent actuator activation and continuous balance control—actual runtime may fall below 3 hours. Battery swapping can mitigate this issue to some extent, but it increases operational costs and system complexity.
By contrast, traditional industrial robots are usually wired for power, enabling 24/7 continuous operation. In factory environments, power supply infrastructure is relatively easy to deploy, making wired operation a non-issue. Humanoid robots, however, derive much of their appeal from mobility and flexibility. Once constrained by power cables, this advantage is significantly diminished.
A Rational Perspective from Mechanical Engineering
For practitioners in mechanical engineering, it is essential to approach the development of humanoid robots with calm and rational judgment. Technological progress is never linear, while capital-market enthusiasm often amplifies expectations and distorts reality.
Humanoid robots have undoubtedly made real progress—but separating genuine engineering breakthroughs from marketing narratives remains critical. Only by doing so can the industry move forward on a solid and sustainable foundation.
Theodore Li