v0.2.0 — real physics, dual-arm collab, offset-tower experiment

First release since v0.1.0. Raises the demo's technical ceiling and brings the classroom materials + landing page up to date.

Highlights

• --real-blocks — colored blocks become genuine Newton XPBD rigid bodies that stack, topple and collide. Grasping is a KINEMATIC↔DYNAMIC toggle (XPBD has no weld constraints): the held block's pose is prescribed from the gripper each frame, flipped back to DYNAMIC on release.
• --collab — two-arm collaborative tower relay. Arm A fetches blocks to a single handoff slot, Arm B stacks them, roles reverse to tear it down, then it loops — yielding instantly to any audience input. A single-slot interlock (plus asymmetric timing) keeps the arms from colliding.
• --experiment — Arm B's offset-tower stability lecture. It stacks three blocks with a growing per-layer offset (0 → 4 → 9 cm) and real XPBD decides when the tower falls. Theory: with 10 cm-half-width cubes the top two layers' CoM sits at 1.5·offset, so it topples at offset > 6.67 cm — the 4 cm round survives, the 9 cm round collapses. A live CoM plumb-line overlay turns amber the instant the criterion flips.
• Physics ~40× cheaper — step() 11.9 ms → 0.30 ms in teleport mode via three levers (fewer iterations/substeps, a teleport-mode collide() skip, precomputed render jitter tables); uncapped headless throughput ~279 fps.

Docs & site

• Defense deck (29 pp) + design report (22 pp) updated to v0.2.0, with an offset-tower topple derivation and a physics-optimization section.
• GitHub Pages landing page: https://hollis36.github.io/newton-vla-demo/
• 250 unit + integration tests, 100% passing on Apple Silicon CPU-only.

See CHANGELOG.md for the full list.

v0.1.0 — Initial public release

A 3-minute classroom demo of embodied AI on a MacBook — no GPU, no cloud.

Highlights

• Three interaction modes in one demo:
  • BALL CATCH — MPC + closed-form IK, 62–82% measured catch rate.
  • TALK TO ARM — typed / spoken natural language → Claude VLA → arm program.
  • GESTURES — wave, point, bow, dance (decorative).
• Hybrid VLA pipeline decouples acting from reasoning:
  • Keyword preflight queues the arm in ~1 ms.
  • claude --print refines in the background (~9 s).
  • Generation counter prevents stale workers clobbering fresh commands.
• Industrial dual-arm mode (--industrial) adds a fixed-base Arm B that
  perpetually shuttles a dedicated workpiece while Arm A serves the audience.
• 214 unit + integration tests, ~102 s wall clock, 100% passing.
• 60.5 fps average on Apple Silicon CPU-only.
• Bilingual keyword fallback (English + 中文).

Documentation

• 📄 Design report — 18 pages, LaTeX
• 🎯 Defense slides — 24 pages, beamer
• 🎤 Rehearsal script — 3-minute on-stage walkthrough

Architecture

5-layer module breakdown (Input → Parser → Control → Physics → Render)
with cross-cutting telemetry. 6612 lines of production code across
19 modules. See README for the full breakdown.

What changed across the 8 development rounds

Phase A → test safety net (41 tests)
Phase B → scene.py split + helper modules
Phase C → robustness & expressiveness
Phase D → 4 new features (gestures, slow-mo, telemetry, AI panel)
Round R → 10 review-driven fixes (2 CRITICAL, 4 HIGH)
Round S → polish + docs sync
Round T → effects + control tests (32)
Round U → catcher tests (18)
Round V → VLA subprocess mocks (17)
Round W → Arm B construction refactor
Round X → Arm B idle loop
Round Y → Arm B autonomous workpiece shuttle + fixed-arm drive bugfix