Visual TL;DR
Fragmented AI ModelsDriver
From the article 4 mentionsThe pursuit of truly intelligent embodied agents has long been hampered by the need to train disparate, specialized models for perception, reasoning, and action.
Inefficiency & LimitsDriver
From the articleThis fragmentation leads to inefficiencies and limits the holistic capabilities of AI systems.
Pelican-Unified 1.0Core
From the article 3 mentionsThe introduction of Pelican-Unified 1.0 marks a significant departure, presenting the first embodied foundation model built on the principle of unification.
Unified VLMCore
single visual-language model maps diverse inputs to shared semantic space
From the article 5 mentionsPelican-Unified 1.0 leverages a single Visual-Language Model (VLM) to serve as a unified understanding and reasoning module.
Chain-of-Thought ReasoningCore
From the article 4 mentionsCrucially, it also performs autoregressive chain-of-thought reasoning, generating task- and action-oriented sequences in a single pass.
Simultaneous OptimizationEffect
From the articleThis unified approach allows for the backpropagation of language, video, and action losses into the shared representation, enabling simultaneous optimization of understanding, reasoning, imagination, and action, rather than relying on isolated expert systems.
SOTA PerformanceOutcome
achieves state-of-the-art performance by integrating perception, reasoning, generation
From the article 2 mentionsContrary to the intuition that unification might lead to diluted capabilities, Pelican-Unified 1.0 demonstrates that this paradigm can preserve and even enhance specialist performance.
Unification EnhancesOutcome
proving unification enhances rather than compromises specialist strengths
From the article 2 mentionsContrary to the intuition that unification might lead to diluted capabilities, Pelican-Unified 1.0 demonstrates that this paradigm can preserve and even enhance specialist performance.