Astribot Unveils Lumo‑2: 20+ Complex Household Tasks Demonstrate Full‑Stack Embodied AI

Astribot announced its second‑generation embodied base model, Lumo‑2, together with more than 20 full‑process household task videos (available at www.astribot.com/research/Lumo2) and a technical report (arXiv:2607.11270). The demo covers a wide spectrum of chores, including collaborative box‑folding with two robots, precise physical interactions such as flipping eggs, weighing grains, brewing coffee, and intricate manipulation like tying bow‑ties and zipping luggage.

The tasks are deliberately grouped by capability dimensions: (1) Collaborative cooperation —tasks that a single robot cannot finish, requiring mutual state awareness; (2) Physical understanding —handling liquids, granular matter, and temperature without relying on fixed trajectories; (3) Temporal reasoning —catching falling balls or placing cups on rotating racks with minimal error tolerance; (4) Long‑chain execution —multi‑step processes such as full coffee preparation where early errors propagate; (5) High‑precision dexterity —operations like tying ribbons, ironing clothes, and precise zipper handling.

Lumo‑2 departs from prior world‑action models (VLA, WAM) by first predicting future world states in a lightweight physics‑based latent space and then generating actions from that prediction. This reverse‑order reasoning reduces computational cost compared with explicit step‑by‑step textual planning and avoids the heavy video‑generation approach of traditional world models.

To bridge the gap between training metrics and real‑world performance, the model employs a three‑stage progressive cross‑modal alignment: (1) aligning actions with latent world dynamics; (2) aligning actions with visual and language modalities; (3) joint training on VLM, video, and robot data. This design yields concrete performance gains: end‑to‑end inference speed is 2.71× faster than standard autoregressive decoding with no loss in accuracy, enabling real‑time closed‑loop control.

Benchmark results show Lumo‑2 achieving the best overall scores across four categories—temporal reasoning, physical understanding, long‑process tasks, and high‑precision manipulation—surpassing Pi0.5 and Fast‑WAM. In embodied VLM tests, Lumo‑2 ranks first on most spatial‑understanding tasks. In generalized pick‑and‑place evaluations, it consistently outperforms baselines, especially on unseen‑instruction + unseen‑object combinations.

The accompanying Agent Philia extends the robot’s capabilities from single‑task execution to long‑term household service. Its modular architecture separates interaction UI, inference models, memory, navigation, and manipulation, allowing independent upgrades. A persistent semantic memory records user preferences, interaction history, and task outcomes, which the system uses for context‑aware planning while still enforcing safety checks. Philia also supports multi‑robot coordination, sharing a common semantic space while each robot retains its own map and runtime environment.

From an industry perspective, Astribot positions itself in the “full‑stack” camp that integrates AI models, embodied OS, and rope‑driven hardware. This contrasts with “brain‑only” approaches that focus on universal models and “hardware‑only” strategies that emphasize mechanical design. The company argues that only the triad of model, OS, and body can deliver scalable, sustainable personal robots for homes.

In summary, Lumo‑2’s extensive real‑world demonstrations, novel latent‑space reasoning, accelerated inference, and the long‑term service‑oriented Agent Philia together constitute a rare, comprehensive showcase of full‑stack embodied AI, highlighting both the progress and the remaining challenges in bringing household robots to market.