Skip to main content

Changelog

Version 2.3.0​

Breaking Changes​

  • Neural Band timestamps renamed: NeuralBandBatch.capture_timestamp_ns is now arrival_timestamp_ns, and the sub-sample capture_timestamp_ns is replaced by wristband_timestamp_ns plus device_timestamp_ns (Optional[int]). No silent fallback — old names raise
  • getT_Device_Cpf() and getT_Cpf_Sensor() take a CpfType (CadBased, SvdBased, Default) instead of a bool useSvd. Not implicitly convertible from bool, so existing call sites must be updated

New Features​

  • Neural Band wristband-to-device clock mapping, exposed as convert_from_wristband_time_to_device_time_ns(), convert_from_device_time_to_wristband_time_ns() and get_neural_band_emg_sample_period_ns(). All return None when the recording cannot relate the two clocks
  • CpfType.Default resolves per generation: CAD on Gen1, SVD on Gen2

Behavior Changes​

  • On Gen2, get_transform_device_cpf() now returns the SVD-aligned CPF by default, matching how gaze is reported elsewhere in the stack. On-device vs MPS eye gaze agreement improves from 44.05 mrad to 8.74 mrad. Gen1 unchanged
  • EMG sub-samples no longer carry synthesized timestamps. wristband_timestamp_ns repeats the packet timestamp; the spread lives in device_timestamp_ns

Bug Fixes​

  • getT_Cpf_Sensor() no longer derives the CPF choice from getCadValue, so it composes with getT_Device_Cpf() again (was off by 2.53 deg on the unit measured)
  • Fixed the open-source CMake build: NeuralBandTimeMapper.cpp was missing from core/data_provider/CMakeLists.txt, and the VRS pin predated vrs::StreamId becoming a literal type (GitHub PR #394)

Improvements​

  • Aria stream identifier constants are now constexpr

Documentation​

  • Gen2 tutorials restructured onto one axis and renumbered: 1-4 read a recording, 5 reads MPS output, 6-8 cover one algorithm each across both on-device and MPS (VIO and trajectory, hand tracking, eye tracking), 9 covers Neural Band sEMG
  • New eye tracking and Neural Band sEMG tutorials; every tutorial now downloads its own sample data and runs end to end on Colab
  • Website tutorial pages removed in favour of linking the notebooks on GitHub
  • Camera frames logged to Rerun are JPEG compressed, fixing dropped frames past the notebook viewer's 1 GiB budget

Version 2.2.0​

New Features​

  • Official support for ML Eye Tracking (ML ET): ML ET is our new machine-learning-based eye-tracking algorithm. It replaces the previous approach and delivers noticeably better gaze quality. projectaria-tools now officially supports ML ET results from both sources — gaze computed on-device and recorded into the VRS eyegaze stream, and gaze produced by Machine Perception Services (MPS). Both are returned as the same EyeGaze structure, so existing code needs no changes: read on-device results with get_eye_gaze_data_by_index() and get_eye_gaze_data_by_time_ns(), and MPS results with mps.read_eyegaze() or MpsDataProvider
  • Neural Band (sEMG) data support: Aria Gen2 recordings paired with a Neural Band now expose an emg stream. Access it through get_neural_band_batch_by_index(), get_neural_band_batch_by_time_ns(), and get_neural_band_batch_configuration(). Each NeuralBandBatch carries its EMG sub-samples, accelerometer and gyroscope readings, channel count, and ADC bit depth
  • EMG calibration: NeuralBandEmgCalibration converts raw ADC counts to physical units via adc_to_volts() and volts_to_adc(), alongside accessors for analog gain, ADC reference voltages, and bit width. Channel values are raw ADC counts, not volts, until converted
  • Eye gaze configuration with field provenance: EyeGazeConfiguration now reports algorithm_name, algorithm_version, nominal_rate_hz, and user_calibrated. Per-field provenance is available through get_single_field_provenance() and get_combined_field_provenance(), each returning SUPPORTED, CALCULATED, HARDCODED, or NOT_PRODUCED — so you can tell a measured value from a derived or placeholder one
  • Battery status stream: Gen2 recordings expose 502-1 battery-status
  • rescale_camera_calibration(): rescales a camera calibration to a different image resolution
  • NVDEC hardware decoding on Linux x86_64: H.265 RGB video decodes on the GPU where CUDA is available, complementing the VideoToolbox path already used on macOS. Only the RGB camera takes the hardware path; mono 4:0:0 streams such as SLAM and eye-tracking cameras are decoded on the CPU by design. has_cuda_support() reports build capability
  • Linux aarch64 wheels: published alongside x86_64 and macOS arm64

Improvements​

  • aria_rerun_viewer gained an EMG panel, and now enables the GPS and Neural Band panels only when the recording actually contains those streams
  • viewer_mps runs properly headless: --rrd_output_path is a documented flag, is mutually exclusive with --web, and no longer attempts to open a viewer window when writing to a file
  • Migrated to Rerun SDK 0.33.0
  • Dependencies are now flat: everything the toolset needs is a base requirement, so pip install projectaria-tools installs the complete package. The [all] extra is no longer defined and has become a no-op
  • Upgraded the C++ standard from C++17 to C++20

Bug Fixes​

  • Fixed a shared library symbol collision that could occur when projectaria-tools and vrs_health_check are installed into the same environment

Behavior Changes​

  • get_time_sync_mode() now returns SubGhz for Gen2 broadcaster and receiver recordings, rather than skipping the mode
  • Python 3.9 is no longer supported. The minimum supported version is Python 3.10; wheels are published for 3.10, 3.11, and 3.12

Version 2.1.4​

New Features​

  • Added support for loading Aria Gen2 eye-tracking CSV output in the MPS reader: parses combined yaw/pitch + per-eye yaw/pitch, back-fills combined_gaze_origin_in_cpf, spatial_gaze_point_in_cpf, and depth from ray vergence
  • Exposed new mps.get_gaze_vergence_point() Python API that computes the vergence point of two eye gaze rays with independent per-eye pitch and arbitrary IPD, with built-in guardrails for near-parallel, divergent, skew, and excessive-depth inputs
  • viewer_mps now uses the back-filled Gen2 spatial gaze point (when available) to render a more accurate 3D gaze arrow, falling back to the yaw/pitch + depth approximation for Gen1 recordings

Bug Fixes​

  • Fixed AlsDataLayout to match the actual Gen2 VRS schema: corrected gain_{red,green,blue,uv,ir,clear} from int32_t to float and exposure_time_us from int32_t to uint32_t (these previously read as 0 for every sample)
  • Fixed bisection_timestamp_search to return the truly closest timestamp index instead of whichever side the binary search landed on; also fixed border guards so an exact match on the first or last timestamp now returns its index rather than None (resolves GitHub issue #100)
  • Fixed an SE3 transform bug in viewer_mps's 3D gaze arrow rendering where the displacement vector was being treated as a point, causing the CPF→device translation to be added twice

Version 2.1.3​

New Features​

  • Added random-access P-frame video decoding: getImageDataByIndex() now supports non-sequential frame access for H.265 video streams. The decoder automatically replays from the preceding keyframe, enabling random-access for training data pipelines and video editors

Improvements​

  • Bumped VRS dependency to include ProximitySensorRecordableClass support

Version 2.1.2​

New Features​

  • Added --show-latency flag to aria_streaming_viewer that displays a TimeSeriesView panel for latency metrics in the Rerun UI
  • Added --blueprint CLI flag to load custom Rerun blueprint (.rbl) files, with automatic preservation on stream reconnection
  • Added robust SVD-based method for computing device-to-CPF transformation

Improvements​

  • Migrated to Rerun SDK 0.26.2: unified rr.set_time() API, renamed archetypes to plural forms, and switched from TCP to gRPC transport
  • Migrated vrs_to_mp4 from moviepy 1.0.3 to 2.x and flattened setup.py dependencies

Version 2.1.1​

Improvements​

  • Exposed XPRS decoder on macOS through the decode_oss_frame Python module by adding NV12 pixel format support for hardware decoding
  • Added rerun_memory_limit variable in AriaDataViewerConfig to allow memory management for the viewer

Version 2.1.0​

New Features​

  • Enabled Google Colab support for all Gen2 Python notebook tutorials — including "Open in Colab" links in the documentation website
  • Added external CI support for the newly-added Gen2 unit-test datasets to improve test coverage
  • Introduced Jacobian computation APIs for camera calibration workflows

Improvements​

  • Enhanced the viewer_mps tool for Gen1 data: now supports rotated, non-rectified images
  • Fixed toJson() serialization issues in SensorCalibration, and added full serialization support for the CameraConfig field

Bug Fixes​

  • Corrected contact-mic sensor labels
  • Fixed the binary name for gen2_mp_csv_exporter
  • Addressed other minor stability issues

Version 2.0.0​

🚀 Major Version Release

We're excited to announce a major update of the library — now supporting both Aria Gen1 and Gen2 VRS data with the same C++ and Python APIs you already know. This release also introduces full access to On-Device Machine Perception (MP) data in Aria Gen2 VRS.

API & Data Compatibility​

  • Unified Interface: The C++ and Python APIs remain unchanged — your existing Gen1 code continues to work seamlessly
  • Extended Capabilities: Added APIs to access On-Device Machine Perception data streams available in Gen2 recordings
  • Cross-Generation Support: A single codebase now handles Gen1 and Gen2 VRS data without modification

New & Enhanced Tools​

New Tools​

  • aria_rerun_viewer: A Python-based visualizer for Aria VRS files, built on Rerun for interactive multi-modal exploration
  • gen2_mp_csv_exporter: A Python tool to export On-Device MP data from VRS to CSV, fully compatible with the Machine Perception Service (MPS) format

Upgraded Tools​

  • vrs_health_check: Completely re-engineered and released as a standalone Python package for easier installation & use

Documentation​

  • New Python Tutorials: Our Python tutorials have been completely rewritten into seven focused guides, covering topics such as:
    • Queued multi-sensor data
    • Device calibration usage
    • Multi-device time alignment
    • ...and more
  • Updated Documentation Websites: We now maintain dedicated documentation portals for each generation:
    • Aria Gen2 Documentation Website — new APIs, tools, and data formats
    • Aria Gen1 Documentation Website — original docs for Gen1 workflows