Changelog

v0.2.1 — 2026-07-27

Added six direction-aligned obstacle modes with terrain-matched planar and turning commands.
Added reproducible Stage 66 continuation tooling with equal left/right knee and step-through weights.
Added directional-obstacle support to the live training viewer.
Fixed manual playback instability by allowing the validated actor to remain pinned while checkpoint metrics and live-training state continue updating.
Fixed compact-atlas terrain selection by mapping trainer and viewer tiles by semantic label instead of numeric column.
Added bidirectional interpolation between genuine consecutive RTX frames, with hard cuts for scene resets and unwarped HUD regions.
Added measured simulation/render rates and separate runtime-derived simulation, render, and display hardware labels.
This patch contains no new policy checkpoint; Stage 66d continues training.

Experimental terrain and live-training update.

Released the Stage 65b checkpoint at iteration 10479 and its terrain-scan branch parent at iteration 10280.
Expanded actor/critic observations to 162/176 values with a 104-sample terrain-height scan and a zero-output checkpoint adapter.
Added 35/65 flat/anchored-box terrain generation with curriculum difficulty.
Added bilateral timing, alternating-knee, step-through, foot-flight, pendulum-roll, inter-foot approach, and late-swing touchdown shaping.
Added cardinal velocity weighting and survival-gated speed curricula through 1.0 m/s.
Added real training-environment state passthrough to the split-GPU viewer.
Added a MANUAL CONTROL / TRAIN LIVE • DR button, saved-state mode switching, live terrain/mass/friction telemetry, and smoothed trainer-state rendering.
Published the final live interface screenshot, Stage 65b metrics plot, resolved configs, TensorBoard history, and machine-readable intermediate evaluations.

First public checkpoint release.