Galbot at WRC: Agile Humanoid Robots Can Dance, But Can They Actually Work?

Galbot's Spotlight Moment at WRC

At the World Robot Conference (WRC), Chinese robotics company Galbot showcased its next-generation agile humanoid robot, drawing widespread attention with its fluid and nimble motion control.

The World Robot Conference is a premier global robotics event organized by the Chinese Institute of Electronics and other institutions. Held annually in Beijing, it brings together top robotics companies, research institutions, and industry experts from around the world. WRC typically features three main components — forums, exhibitions, and robotics competitions — making it a key window into global robotics trends and the development of China's robotics industry. In recent years, as China has dramatically ramped up investment in humanoid robotics, WRC has become the go-to stage for companies to unveil new products and demonstrate their technical prowess.

Galbot is a rising Chinese humanoid robotics startup focused on developing general-purpose humanoid robots. The company has made deep investments in motion control, robot perception, and AI-driven decision-making. Their technical approach emphasizes combining reinforcement learning with traditional control theory to achieve more natural and robust whole-body locomotion. From an engineering perspective, achieving such a high level of dynamic balance and motion control is an extremely challenging technical feat — and Galbot has indeed delivered impressive results.

However, a pointed question kept surfacing in Reddit community discussions: flashy agile movements are certainly eye-catching, but if people are going to actually spend money on a robot, it needs to be able to do real work — not just hop around and dance on a stage for a few minutes.

The Mismatch Between Technical Challenges and Market Needs

This discussion revealed a fascinating phenomenon in the current humanoid robotics industry: the most impressive-looking feats aren't actually the hardest part.

Dancing: Scripted Performance Art

Some community members astutely pointed out that robot dancing looks spectacular precisely because it's fundamentally a highly scripted performance — it doesn't require dynamic adaptation to the environment. Scripted motion means pre-choreographing every joint angle or movement trajectory frame by frame, with the robot executing a predetermined program with precision, much like an industrial robot performing repetitive tasks on an assembly line. This approach can produce very smooth, visually appealing movements in controlled environments, but the robot doesn't need to understand the semantics of its own actions or adjust based on environmental changes. In contrast, autonomous motion requires the robot to perceive its environment through sensors, understand task objectives, and generate motion strategies in real time. The technical gap between the two is enormous: the former is essentially a playback problem, while the latter involves the complete closed loop of perception, planning, decision-making, and control.

This draws an interesting parallel with the development trajectory of large language models (LLMs) — people initially assumed artistic creation would be the hardest domain for AI to crack, yet it ended up being simulated earlier than structured, procedural tasks.

In other words, the motion control that enables a robot to dance gracefully, while extremely difficult from an engineering standpoint, is a relatively closed, pre-definable problem. The real challenge lies in general-purpose task capability in open environments.

From Balance to Tasks: The Ever-Moving Goalpost

One user in the discussion made a thought-provoking observation:

"There was a time when people widely believed that machines could never solve bipedal balance. Now bipedal balance is commonplace, and people don't even marvel at it anymore. Soon the standard will become — 'Sure, they can fold laundry, but unless they can tailor a suit from scratch, I'm not impressed.'"

Bipedal dynamic balance is one of the most fundamental and core technical challenges for humanoid robots. Unlike quadruped or wheeled robots, bipedal robots have an extremely small support area with a high center of gravity, making them inherently unstable. Achieving stable walking requires real-time solutions to complex dynamics equations, involving technologies such as Zero Moment Point (ZMP) control, Model Predictive Control (MPC), and whole-body dynamics optimization. In recent years, the introduction of Reinforcement Learning has enabled robots to learn robust balance policies through millions of trial-and-error iterations in simulation environments, which are then deployed to physical robots through sim-to-real transfer techniques. This has dramatically accelerated the maturation of bipedal balance technology. Boston Dynamics' Atlas, Tesla's Optimus, and several Chinese companies have all made significant progress in this area.

This perfectly illustrates the "moving target" effect in humanoid robot evaluation criteria. Every time a technology once deemed impossible is achieved, public expectations quickly ratchet up, and previous accomplishments are taken for granted.

The Boiling Frog of Technological Progress

Some commenters likened the current pace of technological evolution to the "boiling frog" — we may already be well into the process of a technological singularity, but there's no obvious starting marker.

From early discussions about the Turing test to technology demonstrations on Scientific American programs decades ago — demonstrations that seemed novel at the time but now appear incredibly crude — the cumulative effect of technological progress often only becomes striking in hindsight.

For humanoid robots, the next few years will be a critical window. As the community discussion suggested, once world task models and safety mechanisms mature, humanoid robots will gradually enter people's daily lives. World task models are a cutting-edge concept in robotics and AI, referring to giving robots a general understanding of the physical world — including objects' physical properties, spatial relationships, causal reasoning, and long-horizon task planning capabilities. This concept echoes Yann LeCun's "World Model" vision, emphasizing that AI systems need to build internal representations of how the real world works, rather than merely doing pattern matching on specific tasks. In the robotics domain, world task models mean that a robot can autonomously decompose tasks, plan action sequences, and handle exceptions during execution in entirely new environments, based on natural language instructions or high-level goals. Google's RT series of models, Toyota Research Institute's Diffusion Policy, and the Vision-Language-Action (VLA) models being explored by multiple startups can all be seen as early steps toward world task models.

Marketing Tool or Productivity Tool?

The discussion also surfaced a pragmatic observation: dancing is currently one of the primary actual use cases for these robots — as a marketing tool.

One user bluntly stated that a dancing robot attracts more attention than scantily-clad models at a booth, driving awareness for products. Behind this slightly tongue-in-cheek comment lies an industry reality: current humanoid robots are still in the stage of "demonstrating technical capability" rather than "creating real value."

The Laundry-Folding Bar: The Litmus Test for General Manipulation

Multiple community members independently brought up the seemingly simple task of folding laundry. "Come talk to me when it has the hand dexterity to fold clothes" — this statement captures the core demand of the skeptics.

Folding laundry has become a benchmark for humanoid robots because it requires:

• Perception and manipulation of soft, non-rigid objects
• Dynamic adaptation to uncertain environments
• Fine hand dexterity and force control

Robotic manipulation of deformable objects (such as clothing, ropes, food, etc.) is widely recognized as one of the hardest problems in robotics. Unlike rigid objects, deformable objects have infinite-dimensional degrees of freedom in deformation — their shape continuously changes during manipulation and is difficult to describe accurately with simple mathematical models. This means traditional model-based grasping and manipulation algorithms are nearly impossible to apply directly. Furthermore, perceiving deformable objects is extremely challenging — due to severe self-occlusion and constantly changing shapes, vision systems struggle to accurately estimate their complete 3D state. Current research directions include learning-based manipulation strategies, the use of tactile sensors (such as GelSight and other optical tactile sensors that provide high-resolution contact surface information), and methods that combine large-scale pretrained models with manipulation policies. Despite significant academic progress, reliably folding laundry, making beds, and similar tasks in unstructured home environments remain far from commercialization.

On the hand dexterity front, dexterous hands are the critical end-effectors for humanoid robots to achieve general-purpose manipulation capability. The human hand has approximately 27 degrees of freedom and thousands of tactile receptors, enabling it to perform an enormous range of tasks from precision assembly to soft object manipulation. Designing robotic dexterous hands requires balancing the number of degrees of freedom, actuation methods, sensor density, and cost. Current mainstream dexterous hand designs include direct motor drive, tendon-driven, and hydraulic systems. Force control is one of the core capabilities for dexterous manipulation, requiring the robot to precisely control contact forces with objects rather than merely controlling joint positions. In tasks like folding laundry, the robot must simultaneously coordinate the position and force of multiple fingers, maintaining a stable grasp on the fabric while completing folding motions — placing extremely high demands on sensor precision, control algorithm responsiveness, and hardware reliability.

These are precisely the weakest links in current robot technology, and the gap that must be bridged to move from "performance" to "work."

Conclusion: Beyond Agility — Where Does the Real Value of Humanoid Robots Lie?

Galbot's showcase at WRC undeniably demonstrated the strength of Chinese robotics companies in motion control. But the consensus across the entire community discussion was clear: agility in movement is a necessary condition, not a sufficient one.

What truly determines the commercial prospects of humanoid robots is whether they can reliably perform economically valuable real-world tasks. The industry's next milestone won't be a more dazzling dance routine, but rather those mundane yet difficult everyday tasks — folding laundry, tidying rooms, handling manipulation in complex environments.

When robots no longer need to dance to attract attention, and instead win the market through genuine labor value, the humanoid robotics industry will have truly moved beyond its "performance era."