Changelog & Updates
Chronological progress of the Base Layer Robotics effort.
Implemented Model Context Protocol (MCP) server support, enabling direct agentic interaction with Dojo's training and evaluation pipelines. Hardened CLI payloads, added experiment hooks, and unified logging policies to support large-scale automated research sweeps.
Migrating Dojo towards a single standard where every world exposes raw sensor sources. Entities now own their sensors, brains, and memory, deprecating legacy omniscient observe APIs in favor of embedded-ready SensorSuite pipelines.
Upgraded physics capabilities and implemented a command-conditioned balance controller. The underlying kinematic integrations and pseudo-huber temporal budgets are now fully stabilized for hardware transfer.
Shipped the initial GPU-resident VCB for train-hp, allowing multi-actor clustered ingest with persistent step-stamped snapshots to optimize steady-state learning.