The arrival of DeepSeek as an open-source platform has ignited a flurry of activity within the tech community, drawing numerous collaborators eager to push the boundaries of technologyThis surge has led to intensive exploration in the realm of embodied intelligence, where researchers are now poised to unlock new capabilities—ranging from long-text processing to multimodal reasoningBut amidst this burgeoning landscape, the question looms: who will be the pioneer to fully leverage the potential of DeepSeek?

Those taking on the ambitious challenge of all-modal information processing are standing on the shoulders of this giantWhile humans swiftly navigate the multifaceted world of sensory input, for robots, the task remains fundamentally complexWang Qibin, the founder and CEO of Lingchu Intelligent, notes that humans accomplish simple tasks—like reaching for a remote control—with relative ease, owing to their ability to integrate various modalities including vision, hearing, and touch

Advertisements

In contrast, for a robot, this seemingly simple action is fraught with dependencies that encompass a broader range of movements and decisions.

At the perceptual level, a robot must leverage visual sensors such as cameras for localization and navigation, potentially incorporating depth sensors to glean environmental data that guides its subsequent actionsIn the brain of the machine, tackling the task of retrieving a remote control requires real-time awareness of the environment’s dynamic changes and a continual assessment of its own conditionsFor instance, if a sofa obstructs access to the remote control, the robot's internal algorithms must reconfigure its arm's trajectory or adapt its gripping technique accordingly.

Even for just this one action of obtaining the remote, the complexity mountsThe robot must adeptly control its end-effectors, whether they are grippers or dexterous hands, and modulate their grip strength and tactics based on factors like the remote's size, shape, and weight

Advertisements

"If the remote is smooth, the feedback from the gripping force must be incredibly sensitive to ensure it holds the remote securely without it slipping away," Wang emphasizes.

Humans, equipped with advanced cognitive capabilities, can think and execute such tasks in mere seconds due to the natural multisensory narratives of daily life that inform their understanding and decision-making processesThe interplay of textual, visual, and auditory stimuli enriches the human experience, allowing individuals to comprehend and articulate complex ideas holistically.

According to the team from Peking University, "This flow of all-modal information is crucial for the transition of large models toward general artificial intelligence." They advocate for the expansion into all-modal capabilities as the next breakthrough for the Deep Seek R1 framework“We must construct a closed-loop cognitive system of perception-understanding-inference within complex decision-making scenarios to expand the boundaries of intelligence in the realm of embodied computing.”

As things stand, the Align-DS-V model has already broadened the DeepSeek R1 series to encompass visual-text modalities

Advertisements

The team's aspiration for an all-modal large model—capable of accepting any input and generating any output—foreshadows a new milestone in artificial intelligenceThe common thread remains: aligning these all-modal large models with human intentions poses a significant challenge ahead.

On another front, the paradigm of reinforcement learning (“RL”) continues to show considerable promise, particularly in the evolution of the DeepSeek R1-Zero and Align-DS-V modelsBuilt from the ground up to rely solely on reinforcement learning—eschewing human expert annotations for supervised fine-tuning—DeepSeek R1-Zero exemplifies this approach.

Chen Yuanpei, a co-founder of Lingchu Intelligent and a protégé of the esteemed Li Feifei, articulates the essentiality of reinforcement learning for robots operating within increasingly complex interactive environments

• European Stocks Look for a Turnaround
• India's Central Bank Hints at First Rate Cut in Five Years
• Effective Investment Drives Quality Gains
• South Korea's GDP per capita Surpasses Japan's
• Digital Economy Drives New Consumption

The intricacy of a robot's interaction with its surroundings is challenging to codify into a precise model manuallySolely relying on deep learning can render a robot's flexibility in performing diverse tasks under varying scenarios precarious, especially given the need for abundant, quality training data—a process fraught with high costs.

He elaborates that employing reinforcement learning can facilitate training robots by altering reward structures, allowing them to utilize extensive simulated data for their education in practical environmentsMost current robotic entities on the commercial market demonstrate single-dimensional gripping capabilitiesHowever, actual operational conditions rarely present objects in isolation, as robots frequently encounter cluttered environments where objects frequently overlap and obstruct one anotherConsequently, enhancing a robot's multi-skill coordination is key.

“A robot's ability to quickly comprehend the complex physical properties of items is imperative,” Chen asserts

A pragmatic pathway could see robots achieving object generalization within a limited skill set over the next few yearsThis could mean, for example, that within packaging domains, robots adeptly sort and pack varying items while scanning barcodes—reflecting ongoing goals for iterative advancements.

As competition intensifies within the field of embodied intelligence, the direction increasingly zeros in on specific applicationsLast year, Lingchu Intelligent unveiled the Psi R0, an end-to-end embodied model grounded in reinforcement learningThis innovative model enables dexterous hands to link multiple skills synergistically and generate intelligent, reasoning-capable agents through mixed training, achieving hybrid operations across different environments.

In parallel, the Epoch Era launched the ERA-42, a foundational large-scale robotic model, showcasing operational capabilities when integrated with their XHAND1 dexterous hands

Demonstrations highlighted their system's proficiency in tasks such as hammering in nails or drilling screwsFurthermore, on January 9, Galactic Universal introduced GraspVLA, touted as the first fully end-to-end embodied grasping foundational model globallyIt leverages synthetic data during pre-training to adapt foundational capabilities to specific scenarios using minimal sample learning.

A clear trend emerges from the latest launches: companies are increasingly intertwining vast models with practical operational contextsThis begs the question: does this indicate a convergence of embodied intelligence applications as competencies become more closely aligned? While large models empower robots with advanced learning, semantic comprehension, reasoning, and decision-making skills, the transitions from understanding to execution also engage various algorithms and necessitate cohesive software-hardware collaborations.

Rather than merely depicting a convergence of applications, it's safer to assert that businesses are adopting a more pragmatic stance

Firms are set to prioritize operational scenarios, steadily refining their robots' skill levels while enhancing software-hardware interconnectivityThe landscape of large-scale embodied intelligence development has only just begun, and companies identifying clearer applications and capabilities will likely find greater cost-effectiveness as a result.

As exemplified by models like Align-R1-V, the emergence of cognitive frameworks that are capable of cross-modal perception enhances a robot's ability to comprehend multimodal inputsHowever, achieving this necessitates the addition of action generation modules, real-time controlling systems, interaction data frameworks, and safety structures to ensure a smooth transition from multimodal understanding to embodied intelligenceIntegrating software models with hardware systems—be it robotic arms, dexterous hands, or drive chips—requires time and concerted effort.

With the sensational rise of DeepSeek, the shift from textual modes to multimodal and all-modal landscapes opens up an array of new challenges