Striding AI Unveils Physical AI Solution for 24/7 Retail at APRCE 2026

SINGAPORE, October 23, 2026--On October 8, 2026, Striding AI unveiled its Physical AI solution for retail, including two self-developed robots and a robot management platform, at the 22nd Asia-Pacific Retailers Convention & Exhibition (APRCE 2026).

Designed for human-robot collaboration in real-world retail environments, the solution automates carrying, shelf replenishment, inventory checks, and patrols, allowing employees to focus on customer service. Robots can also undertake repetitive overnight work, reducing the need for late-night employee shifts while supporting 24/7 convenience-store operations. Retail presents particular challenges for robotics, including extensive product assortments, dense shelving, narrow aisles, changing merchandising conditions, and frequent customer interactions.

Under reported deployment conditions, the system completes up to 99% of tasks autonomously, covering task assignment, perception, navigation, object recognition, grasping, and execution, with human intervention when needed.

"Retail is one of the most complex and dynamic environments for robots to operate in, with diverse products, changing workflows, and frequent human interaction. By bringing Physical AI into real retail environments, we can continuously learn from deployment and turn real-world experience into more capable robotic systems," said Song Yao, Founder and CEO of Striding AI, during his keynote, "Physical AI and the Future of 24/7 Retail."

An Integrated "Embodied Brain" for Physical AI That Continuously Learns from Real-World Tasks

Striding AI's system-first approach combines foundation models, robotic perception and control, real-world action data, robot hardware, and deployment infrastructure. Its "Embodied Brain" integrates world-action models, reinforcement learning, embodied agents, and agentic infrastructure to help robots to perceive, reason, act and improve through real-world interaction.

Its SLM-0.5 (Latent World Action Model) generates actions in latent space rather than through repeated pixel-level prediction, reducing computing requirements and latency. Under reported evaluation settings, it achieved 98.6% success on LIBERO and 91.22% on RoboTwin, with 187 ms latency per action chunk - up to 24 times lower than comparable pixel-space world-action models.

STEAM uses reinforcement learning and human feedback in an execution-feedback-learning-redeployment loop, allowing robot policies to improve from actual operations rather than relying only on pre-collected training data.