Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Fixed-horizon CusADi decode artifacts and generated CUDA source for 50, 80, 100, 150, and 200 sample trajectories.
dhb_xr-build-cusadi-decode for compiling supported CusADi decode libraries into a dhb_xr-owned cache.
GPU decode documentation covering install paths, cache behavior, runtime fallback policy, and deployment troubleshooting.

generate_trajectory(..., backend="auto") documentation now presents FATROP as the preferred single-trajectory backend with explicit fallback policy.
The cusadi optional extra now depends on CasADi and PyTorch rather than an external editable CusADi checkout.

Torch DHB-QR decoder path.
VLA action-head, chunk rollout, SmolVLA adapter, and waypoint controller utilities.
Focused tests for action-head behavior, chunk rollout continuity, VLA pipeline stages, LIBERO orientation handling, Torch DHB-QR decoding, OAT tokenizer diagnostics, waypoint control, and solver retargeting.

Improved VLA pipeline token streaming and OAT prefix decoding behavior.
Updated CasADi/Fatrop retargeting paths and timing documentation.
Standardized VLA dataset adapter outputs on DHB-XR's public wxyz quaternion convention.
Removed unused dtaidistance from database extras and included einops in the all extra.

LIBERO quaternion/action semantics and rollout orientation handling.
Fatrop boundary handling and trajectory-derived initial frame usage in retargeting.

FAST Tokenizer: Frequency-space action tokenizer using DCT + BPE compression, no training required (~28x compression on invariant sequences)
OAT-style Tokenizer: Ordered Action Tokenization with Register Encoder, Finite Scalar Quantization (FSQ), and Masked Nested Dropout for anytime prefix decoding
FSQ Module: Finite Scalar Quantization with deterministic rounding -- eliminates VQ codebook collapse by construction
Register Encoder: Transformer-based temporal compression using learnable register tokens with causal-last attention
Nested Dropout: Masked nested dropout module for inducing coarse-to-fine token ordering (Matryoshka-style)
OAT Decoder: Cross-attention Transformer decoder for reconstructing sequences from register token latents
Benchmark Suite: Three benchmark scripts comparing all 7 tokenization methods on rate-distortion, SE(3) stability, and prefix decoding
Interactive Notebook: Comprehensive tutorial notebook (tutorial_action_tokenization.ipynb) with head-to-head comparisons and visualizations
Manuscript Updates: Expanded Section 7 with FAST/OAT descriptions, new Appendix with empirical comparison tables

SE(3) Invariance: Fixed invariant token stability under SE(3) transforms by using trajectory-relative initial frames (init_pose from trajectory start) instead of fixed global axes. This achieves exact invariance (0% token change, MSE = O(1e-13)) versus the previous approximate invariance
BPE Token Comparison: Improved token stability reporting to account for different-length BPE sequences and separate pre-BPE (DCT integer) from post-BPE token differences

einops dependency: Added to [tokenization] optional dependencies for FSQ implementation

VLA Integration: Comprehensive LIBERO and RoboCASA support for Vision-Language-Action benchmarks
LIBERO-PRO Support: Full integration with LIBERO-PRO perturbation robustness testing
Trajectory Adaptation: Fatrop and CasADi solvers for trajectory retargeting (decode ~5-10ms, Fatrop NLP <1s)
Swap Demo: Compelling demonstration showing DHB-XR vs naive replay under spatial perturbations (6.5cm improvement)
Simulation Integration: Real-time LIBERO environment execution with OpenCV visualization and video recording
VLA Documentation: Extensive documentation explaining how DHB-XR addresses VLA spatial generalization challenges

Repository URLs: Updated all links to point to robodreamer organization
Documentation Updates: Added VLA context throughout README and documentation

Solver Integration: Proper Fatrop/CasADi solver selection in trajectory adaptation functions

MkDocs Documentation: Migrated from Sphinx to MkDocs with Material theme
GitHub Actions CI/CD: Automated documentation builds and deployment
Encoding Method Enum: Replaced string literals with type-safe EncodingMethod enum
GitHub Pages: Documentation hosting setup

Trajectory Preprocessing: Robustness modules for handling reversals and zero-motion segments
DHB-Token: VQ-VAE tokenization pipeline for discrete action representations
Generative Models: Variational Flow Matching (VFM) for multi-modal trajectory generation
Motion Database: Similarity search and retrieval with DTW support

Core DHB-XR Implementation: DHB-DR (Double-Reflection) and DHB-QR (Quaternion-Relative) invariant encoding
Time-Invariant Reparameterization: DHB-TI for speed-independent trajectory representations
GPU Acceleration: PyTorch-based batched operations and CusADi for large-scale optimization
Fatrop Integration: Fast structured OCP solver for trajectory optimization
CasADi Support: General nonlinear optimization for trajectory adaptation
Visualization: Enhanced SE(3) trajectory plotting with matplotlib
Jupyter Notebooks: Comprehensive demo notebook with interactive examples

Dependencies: Added CUDA support with PyTorch 2.0+ and optional GPU features
Documentation: Enhanced README with performance benchmarks and usage examples

Initial Project Structure: Basic DHB encoding/decoding functionality
Project Configuration: Pixi environment, GitHub Actions, and development tooling
Basic Testing: Initial test suite setup
Documentation: Sphinx-based documentation framework