To help external partners clearly understand what HanabiAI can already provide at the compute-platform level and what we are currently validating, we have, for the first time, systematically organized and published the positioning and capabilities of the Embodied AI Compute Platform on our official website.

This Changelog highlights two key points:

Clarifying the capabilities that the Embodied AI Compute Platform can already provide for general HPC & AI scenarios.

Outlining the directions currently being tested and validated for embodied-intelligence scenarios, enabling research institutions and partners to align expectations with us.

1. Established Capabilities for General HPC & AI Workloads

For general HPC and AI workloads, the Embodied AI Compute Platform currently provides:

• Unified job submission and core user experience whether it's simulation, numerical computation, or AI training/inference, tasks can be submitted, monitored, and retrieved through a unified workflow.

• Centralized view across multiple clusters and partitions supports integration with multiple clusters, heterogeneous partitions, and hybrid local ＆ cloud setups, allowing users to view resource usage and queue status from a single interface.

• Foundational observability and log aggregation aggregates task status, resource utilization, and key logs in one place, reducing the cost of debugging issues and reproducing experiments.

These capabilities have been iterated and battle-tested extensively within our internal environment, and we consider them production-ready for real-world projects.

2. Capabilities Under Testing and Validation for Embodied-Intelligence Scenarios

In the embodied-intelligence domain, we position the Embodied AI Compute Platform as the cluster-side engine that supports the full closed loop of simulation → training → evaluation → replay. At this stage, these capabilities are in the design, implementation, and testing/validation phases, primarily including:

• Workflow design for large-scale parallel simulation and scenario management. How to more effectively manage embodied-intelligence simulation tasks, scenario versions, and datasets.

• Orchestration and tracking for reinforcement learning / imitation learning training. How to reliably operate large-scale RL/IL training on the Embodied AI Compute Platform, while ensuring experiment versions, parameters, and results remain traceable and comparable.

• Support for policy evaluation and Sim-to-Real validation. How to better organize policy evaluation, replay, and transfer verification, providing meaningful feedback data for on-device robots.

These directions are currently being explored through small-scale pilots and PoC (Proof-of-Concept) collaborations within our internal environment and with selected partner organizations. We will continue refining and tightening these capabilities over time.

3. Relationship Between the Embodied AI Compute Platform and the Embodied AI SoC

In this update, we also clarified the relationship between the Embodied AI Compute Platform and the Embodied AI SoC:

• The Embodied AI Compute Platform runs on the cluster / data-center side, handling simulation, training, evaluation, and model engineering.

• The Embodied AI SoC runs on the robot device side, responsible for multimodal inference, policy execution, and safety-critical loops.

• The two are connected through “model export and on-device adaptation” and “runtime data and metric feedback”, forming a closed loop.

This means that, in embodied-intelligence scenarios, the role of the Embodied AI Compute Platform is essentially “the cluster-side workbench that prepares and iterates models for the on-device robot brain.”

Going forward, we will continue to record tangible progress along both tracks—general HPC & AI scenarios and embodied-intelligence scenarios—in future Changelog updates.