Current project status

The current build is recorded at 650 g with a 2S battery, Mini560 converter and USB-C charger. Freestyle now runs a finite program once and uses actual body events for recovery. The October 27 model is activated on the workstation; a separate supported stretch verified all nine targets. Release choices, general stretch calibration, loaded motion and sustained Full Auto initiative remain unfinished.

Physical build

650 g; separate battery, converter and charger The October 26 hardware record matches the physical build to the Blender assembly. New exports include a one-piece power carrier and a three-part training stand. Pack capacity, runtime and printed stand fit remain to record.

Motion training

Finite-event contrast model activated October 27 The 339-example continuation matches 40 of 62 frozen references. Programs no longer request their own continuation. Simulated idle, jam and tip cases complete, but some release choices and wave counts remain wrong; physical recovery is not yet demonstrated.

Measured execution

Arrival feedback replaces timing alone Target-and-speed commands use the fourth-wire shaft readings. Supported Stand and one Walk cycle completed after the stack repair. Loaded chatter, floor walking and generated recovery remain unresolved; the 0.8.127 source also moves boot/rest requests to Freestyle.

Reviewed gestures

Supported stretch verifies nine movement targets The October 27 stretch trial passed after calibration positioning, with no adjusted targets. Its asymmetric crank compensation is retained as trial evidence, not a generic training motion. Earlier brisk-lift and sway corrections retain their separate reviews.

MetaHuman continuity

Existing intentions, selected memory and actual outcomes Full Auto can address existing work and choose when to reassess. Live retests completed status queries and drained the queue, but did not select self-directed conversation or motion.

Runtime and models

Idle conversation context and on-demand weights Unanswered questions retain context without appearing as stuck processing. Queue cancellation and snapshot publication have repairs; llama.cpp loads models on demand and defaults to releasing idle weights after 300 seconds.