As we dive into the technological landscape of 2025,a significant development is making waves: DeepSeek (R1).This innovation is,without a doubt,set to transform the narrative surrounding humanoid robots,a sector in which advancement has been sluggish,primarily due to prohibitive costs and the reliance on scarce data for development.

In recent months,the tech industry from China to the international stage has taken notice of DeepSeek.Numerous tech companies have announced their engagement with or development of products based on the DeepSeek platform.While some of these endeavors are driven by marketing trends,many recognize the groundbreaking potential of DeepSeek R1.This includes companies that may superficially seem unrelated,such as humanoid robotics manufacturers.

For instance,on February 7,Ubtech,the first publicly traded humanoid robotics company in China,revealed that it is currently testing the efficacy of DeepSeek technology in humanoid robot applications.They are exploring areas such as multimodal human-robot interaction,command comprehension in complex environments,and task decomposition and planning in industrial settings.Ubtech expressed optimism that the powerful reasoning capabilities of this advanced model will help tackle intricate challenges,bringing humanoid robots closer to human-like thought processes and behavior.

But Ubtech isn’t alone in this revelation.Just days prior,Brett Adcock,the founder of the Silicon Valley humanoid robotics company Figure,publicly declared on the X platform that the company would terminate its collaboration with OpenAI to shift towards an internal,end-to-end robot AI development.Adcock hinted at a major breakthrough and plans to unveil unprecedented advancements in humanoid robotics within the next 30 days.This pivot has been interpreted by many as a strategic move to develop AI using the DeepSeek R1 model,among other open-source frameworks.

Even other companies focused on delivering cost-effective humanoid robots,such as Yushu Technology,have reported establishing deep collaborations with DeepSeek.This wave of interest among humanoid robotics firms indicates a significant re-evaluation of development routes inspired by DeepSeek’s revolutionary model,much like that of Ubtech.

DeepSeek is,thus,poised to reshape the trajectory of humanoid robotics,tackling the key obstacles that have hampered widespread adoption.The pivotal issue is cost,especially surrounding the immense computational power required to train robots with embodied intelligence—a substantial barrier for many startups in this field.

The financial burden associated with high-powered computation can deter many emerging companies,especially when those with considerable resources,such as tech giants,dominate the AI landscape thanks to their financial backing.For most humanoid robotics startups,the prohibitive costs of computational resources present an insurmountable hurdle,mirroring the issues that plagued the industry as a whole,such as a lack of robust data.

This data scarcity was notably highlighted when OpenAI disbanded its humanoid robotics team four years ago due to an insistent lack of data.Thus,when the company Zhiyuan Robotics introduced a million-strong real-machine dataset called AgiBot World last year,it captured the attention of many in the field as a potential remedy to this widespread challenge.Yet,even this dataset was seen by some as merely a drop in the ocean,insufficient for achieving the dream of embodied intelligence.

Imagine a talented athlete denied access to training facilities and equipment; without the right environment,talent may go unrecognized.However,DeepSeek R1 offers a potential solution to alter this narrative radically.

The model’s pricing structure highlights its advantages clearly: 4 RMB per million tokens for cache misses,1 RMB per million for cache hits,and 16 RMB per million tokens for outputs.These figures starkly contrast with OpenAI’s defined API pricing,which ranges from 55 RMB to 438 RMB.Such a shift equips humanoid robotics companies with an opportunity to free themselves from the shackles of exorbitant computational expenses and invest more directly in robot development,facilitating rapid innovation and product upgrades.

Moreover,what sets DeepSeek R1 apart is that it delivers state-of-the-art performance capabilities comparable to OpenAI’s top-tier models,enabling powerful reasoning in mathematics,coding,and natural language.Ubtech’s ambitions could be realized as they enhance their humanoid robots' command comprehension,task planning,and execution capabilities,aiming for a closer approximation to human cognitive and behavioral patterns.

Additionally,DeepSeek R1 boasts algorithmic improvements,positioning the model to operate effectively with less data than previously necessary.Through data distillation techniques,it can automatically prioritize high-value data and leverage adversarial training to generate synthetic data,dramatically reducing the cost of high-quality code data acquisition from 0.8 RMB to as low as 0.12 RMB for every 100 tokens.

In an official announcement,DeepSeek highlighted that R1 employs reinforcement learning in the post-training phase,which significantly boosts the model's reasoning abilities,even with minimal labeled data.This is of paramount importance for a sector afflicted by data limitations,offering potential guidance for the future of embodied intelligence.

With the model being open-source,developers and manufacturers are now permitted to modify and refine it,providing them the flexibility to adapt DeepSeek-R1 to meet their specific needs effectively.This is a revolutionary feature,enabling all humanoid robot entities to cultivate top-tier models or even redefine their approach towards embodied intelligence,significantly lowering the barriers for entry as they harness DeepSeek’s capabilities.

Beyond just DeepSeek R1,another game-changer has emerged: DeepSeek-VL,a new visual-language model recently announced by the team.While only the 1.3B and 7B model versions have been revealed,DeepSeek-VL promises leading performance based on real-world scenarios while maintaining robust language capabilities—a quality that many other large models seem to overlook.

The team's research emphasizes that during training,they incorporated extensive multimodal and language data into their model’s development.Nevertheless,the real test of this visionary approach will unfold in practical applications over time.For humanoid robotics manufacturers,DeepSeek-VL could signify the next pivotal advancement,complementing DeepSeek-R1 and significantly accelerating their path to practical deployment.

In essence,as DeepSeek rises,so too does the possibility for humanoid robots to evolve into indispensable tools across industries,heralding a new era where human-like interfaces and interactions are seamlessly integrated into everyday experiences.