2025 was an insane year for robotics research

Long time model architecture/training challenges were solved and major progress was made on data collection techniques, understanding data quality, and data recipe. This gives Physical AI companies the confidence to finally start investing in large scale data collection.

You saw companies like Figure, Dyna, and PI reach >99% success rates in real life deployments in diverse settings by leveraging RL innovations. Many frameworks were developed for self improving and self-recovering robot models. Researchers figured out how to prevent overfitting in VLA fine-tuning while retaining generalist capabilities. Which means we can build toward generalist models by merging specialist models.

Robots can also move much more agilely from methods like action chunking and FAST tokenization. We see robots able to exhibit smooth full body control at human speeds and not slow or choppy movement.

Roboticists showed how to effectively fuse multi-modal sensor data for huge policy improvements. Integrating vision, language and tactile data was challenging, but doing so opens the door for many contact rich tasks that require a granular sense of force. Force awareness also solves for common issues like visual occlusions.

System 1/2 architectures were hardened to handle long-horizon planning/ task orchestration which enable robots to perform jobs that require series of tasks. Gemini Robotics-ER 1.5 introduced Chain-of-Thought reasoning to physical agents, allowing them to parse constraints and evaluate Semantic Safety.

Memory advancements allowed robots to maintain long-term spatio-temporal reasoning, breaking the "memory wall" with brain-inspired algorithms. NVIDIA's ReMEmber used memory-based navigation, while Titans + MIRAS enabled test-time memorization for sustained performance.

Advancements in the foundation model space also continue to compound progress in Robotics. Better VLMs means VLAs with better spatial understanding and data labeling and processing pipelines that can massively increase in throughput. World models are starting to show promise in data augmentation and policy evaluations.

2025 gave us a small taste of what data scale could do. A glimpse into the future with robots exhibiting emergent intelligence such as zero-shot affordance mapping, visual force sensitivity, and all sorts of general physical reasoning

2026 we get to experience physical AI with 100x the data scale

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。