The Untold Story of Lumo-2 | Astribot R&D Team Deep Dive

In this talk

1. 1:20

   From Lumo-1 to a world model

   Jannan describes Lumo-1 as an early attempt to give a robot rules for tasks such as responding to someone who says they are hungry. For Lumo-2, the team added a world model to help connect vision and action and let the robot find its own approach. Jannan says this can help it move beyond the limits of examples in its training data and reason faster.

2. 4:36

   Testing difficult household tasks

   The team designed tasks around long sequences and precise manipulation, including making coffee, pressing coffee grounds, tying a bow, opening a tea bag, and packing clothes. Jannan and colleagues point out that ordinary household tasks can become difficult when an expected object is missing or soft items block a bag opening. They describe repeated attempts and human correction data as part of improving the robot in real settings.

3. 8:16

   Building an agent and a resilient team

   The team discusses Philia, an agent created to support interaction between people and robots. Philia was given daily self-reflection prompts, and the name came from its own conversation with a team member. Jannan says progress through robotics problems is motivating, while colleagues describe a team culture of calmly analyzing problems and persisting through setbacks.

4. 12:57

   Making robots safer at home

   Asked how people could feel safe living alongside a powerful robot, Jannan describes both behavioral safeguards and compliant hardware. A colleague explains that the robot uses a tendon-driven structure with a non-rigid connection to its motors, so its end-effector force feels gentle and a person can move it by hand while it continues reasoning.

5. 15:04

   Designing tasks around real homes

   A researcher describes the cocktail-making demonstration as a sequence of bringing over a shaker, pouring different liquors, and shaking the mixture. The task requires the model to track earlier steps and context. The team says it designs tasks to resemble everyday home scenarios.