Ropedia Super Week Showcases Six New Research Projects Highlighting Approaches to Robot Learning and Control

Ropedia Super Week features two Ropedia Foundry projects alongside four studies exploring spatial reasoning, world modeling, and robot learning and control

SINGAPORE, October 20, 2026 /EINPresswire.com/ — Ropedia, a Physical AI startup building data infrastructure for robotics and embodied AI, today showcased six research projects as part of Ropedia Super Week. The research contributes to Ropedia’s broader effort to build a data infrastructure layer for Physical AI, connecting structured records of human experience with the capabilities needed to develop more capable robot systems.

Ropedia Foundry’s EgoXperience and HoliWorld projects explore how egocentric human experience can be structured into data for Physical AI research. Four further projects investigate complementary foundations for downstream intelligence, spanning spatial reasoning, physical prediction, and robot learning and control.

Ropedia Foundry: Structuring Egocentric Human Experience
With two new research projects, Ropedia is taking its Foundry data engine further, advancing how first-person human experience is structured into data for Physical AI research.

EgoXperience extracts aligned information about camera movement, depth, hands, and objects from first-person video. Its EgoX-Data-6M dataset contains approximately six million frames, equivalent to around 50 hours of video, providing a resource for embodied AI research.

HoliWorld converts first-person video into structured 3D scene understanding, tracking objects and their interactions over time. Its dataset contains approximately 78,000 question-and-answer pairs across 444 videos and around 100 scenes. On the EgoDynamic4D benchmark, it improved Token F1 from 18.29% to 66.39%.

Spatial Reasoning
The new research pushes AI beyond object recognition, advancing its ability to understand spatial relationships and reason about physical environments.

S-Agent 2.0 answers 3D questions about direction, distance, movement, and route planning from images and video. Its perception toolkit combines visual understanding, precise measurements, and persistent scene memory to help AI agents use the evidence they collect. Without additional training, it improved Qwen3.8-27B’s score by 13.9 points on MMSI-Bench, reaching 66.5, and by 20.5 points on ReVSI, reaching 82.3. It outperformed existing agentic systems across all five spatial reasoning benchmarks evaluated.

World Modeling and Prediction
Building on spatial understanding, Ropedia is exploring how AI can predict physical outcomes and anticipate the consequences of actions.

Ropedia-W0 (research codename: Xperience-0) is a world model that generates first-person video from human or robot motion instructions. The project explores how human and robot data can support learning across humans, robotic grippers, and humanoid robots. In reported experiments, simulated policy scores achieved a 0.996 correlation with real-robot results, while model-guided action selection improved task success to 55%, compared with 26% for random selection and 45% for the best fixed policy.

Robot Learning and Control
Moving from prediction to execution, the research explores how robots can learn skills, carry out tasks, and adapt their movements to different situations.

DONIC develops a generalist controller for dexterous, two-handed robot manipulation, combining human demonstrations with reinforcement and imitation learning. In simulation, it achieved success rates of 88.68% on unseen movement trajectories and 59.76% on unseen objects without fine-tuning.

Ropedia-A0 (Code as Policy) uses coding agents to develop and refine reusable robot-control programs in simulation. The resulting code executes tasks without requiring AI model calls at runtime. It achieved a 70.24% success rate across 42 two-handed robot tasks in the RoboDojo simulation benchmark.

Together, the six projects contribute to Ropedia’s work on the data infrastructure layer for Physical AI. Structuring egocentric human experience provides a foundation for learning from the physical world, while spatial reasoning, world modeling, and robot learning and control address complementary capabilities needed to interpret environments, anticipate outcomes, and execute tasks. Ropedia’s objective is to connect human experience with these capabilities, helping researchers and developers build more adaptable Physical AI systems.

“Physical AI needs more than capable models. It needs data that reflects the physical world and ways to turn human experience into useful resources for research,” said Zhaoxi Chen, Co-Founder and CEO of Ropedia. “Through Ropedia Foundry, we are building infrastructure to structure egocentric human experience. Our other research explores complementary foundations for downstream intelligence, including spatial reasoning, physical prediction, and robot learning and control. Ultimately, these capabilities need to work together to help robots learn from the physical world and act more effectively within it.”

Ropedia will continue developing its data infrastructure and related research, evaluating how these approaches can support learning and performance across a wider range of robot tasks and environments.

Aniket Gupta
Mahou Consulting

Media gallery