Embodied ICL Startups Emerge as Context Scaling Becomes New Frontier
Embodied AI startups are now leveraging In-Context Learning to let robots learn new tasks from single demonstrations, marking a shift toward context scaling as the industry moves beyond data-heavy training.
A week before S1's release, Generalist AI unveiled GEN-1.5, also centered on one-shot learning. The robot can learn a new task in seconds from a single demonstration, with no gradient updates or fine-tuning, and supports human-to-robot transfer, compositional generalization, and sim-to-real migration. Both developments attempt to make robots leverage longer, multimodal context.
In-context learning (ICL) was introduced by OpenAI in 2020 with GPT-3, where the model performs a new task during a single forward pass based on a few examples in the prompt—no parameter updates involved. Over the past three years, context windows expanded from 4,000 tokens to one million, enabling models to process entire books or code repositories, with ICL a major engine behind that growth. Unlike LLMs, embodied models face composite contexts of visual observations, language instructions, and action sequences. Real tasks are often non-Markovian: what to do next depends on what happened minutes earlier. Robot contexts can involve dozens of visual frames per second plus proprioceptive states and actions, holding orders of magnitude more information.
It was not until July 16 that a systematic effort, RoboTTT, proposed by a team including Fei-Fei Li, Jim Fan, and Yuke Zhu, first moved the idea of context scaling into robot visuomotor policies. In August, Generalist and Skild published their follow-ups. Although both companies only released demonstration results without full technical details, discussion of ICL in embodied AI quickly intensified. Whether this path can truly succeed still hinges on many unresolved technical problems.
Against that backdrop, COCO Matrix, a Chinese startup founded in April 2026, is working to make in-context learning the underlying paradigm for embodied intelligence. Founder and CEO Gao Yuxiang, 30, a Xi'an Jiaotong University gifted youth class alumnus who dropped out of a Johns Hopkins PhD program in human-computer interaction, told a tech outlet that when GPT-3 first showed in-context learning, he was already thinking about whether such ability could be transferred to robots. After leaving academia, he worked at Fourier, where he led the full pipeline from teleoperation and data collection to model deployment on full-size humanoid robots. From May to August 2025, he trained a world model with certain generalization capabilities on a full-size bipedal humanoid for a cost of over one million yuan.
Unlike Generalist and Skild, COCO Matrix proposes to move post-training ICL into the pre-training stage. In a recent interview, Gao said that for a single task, collecting around 200 demonstrations is enough to fit it well, and adapting a new task from demonstration to model deployment takes about 18 hours. But even with more data, the widely expected general reasoning ability has not emerged. He also noted that embodied data itself is hard to scale: teleoperation data lacks uniform quality and diversity.
Gao said the industry is now shifting focus from how much data the model has seen during training to whether a robot can quickly learn in a novel environment or task from current context. He cited RoboTTT as an attempt to encode and compress longer histories, while Generalist and Skild demonstrated one-shot adaptation without full fine-tuning. The scaling dimension, he argued, is expanding beyond model size and data volume to include continued learning after deployment.
On the technical challenge of embodied ICL, Gao identified three core difficulties: visual understanding that must combine high-level semantics with precise spatial information; multimodal integration where touch, hearing, and proprioception have far less data than vision and language; and long-context memory that requires compressing, filtering, and retaining relevant historical information over time.
Regarding demonstration length, Gao said that total duration matters less than information density. A ten-minute video of repetitive actions is far less useful than a complete task demonstration. He also emphasized that long context does not mean simply extending a fixed window. Real robot tasks may only need a few minutes of relevant context. COCO Matrix sees long context more as a streaming mechanism: as new context flows in, the model continuously filters and compresses history, retaining useful information from minutes or even an hour earlier. He mentioned that solutions being tested include improving the model's own understanding to decide which history is important, and compressing the large KV cache produced over long sequences using techniques from large language models and video generation, such as linear attention, sparse attention, or routing-based key information selection.
On data, Gao said ICL reduces dependence on exhaustive task coverage but does not remove the need for basic data quantity and diversity. He identified a clear gap in human-robot teaching, real-time correction, and collaborative operation data. The industry lacks mature definitions and collection standards for such interaction data—for instance, when a human corrects a robot during teleoperation, which moments should be recorded and how should actions and feedback be aligned? He said COCO Matrix plans to define human-robot teaching interaction data as a priority.