Release Notes: Robotics AI Suite 2026.2#

September 07, 2026

Autonomous Mobile Robot#

Improved

  • Added opt-in per-stage latency profiling (lock-free ring buffer with a dedicated writer thread, using CLOCK_MONOTONIC to stay immune to PTP clock steps) behind CMake build options, off by default to match upstream.

  • Hardened the ported SLAM sources with mutex/locking corrections around shared LiDAR/IMU buffers, division-by-zero and out-of-bounds guards, main() try/catch wrapping, and removal of an internal-lab IP address from a configuration comment (BDBA information-leakage finding).

  • ros-kpi: Streamlined the benchmarking component and its packaging.

Humanoid Toolkit#

This release extends humanoid support to the Intel® Core™ Ultra Series 3 “Panther Lake” (PTL) platform, introduces NVIDIA Isaac-GR00T foundation-model pipelines accelerated with the OpenVINO™ toolkit, and adds three new LiDAR-based SLAM back-ends for mobile robot localization and mapping, each shipped as a pristine upstream git submodule with Intel changes applied as patches, so they can be evaluated as alternative back-ends without forking the reference navigation stack. All three are validated by replaying public datasets and include reference Intel® Core™ Ultra “Panther Lake” (PTL) core-pinning and frequency-locking scripts plus opt-in per-stage latency profiling.

New

  • LIO SLAM: FAST-LIO2 — Added a computationally efficient, tightly-coupled LiDAR-Inertial Odometry pipeline (fast-lio2-demo) ported to ROS 2 Jazzy on Ubuntu 24.04. Includes new LiDAR configurations, a configurable OpenMP thread count sized from runtime CPU affinity, a preprocess crash fix for Velodyne scans missing a time field, and validation via public NCLT dataset replay.

  • LIVO SLAM: FAST-LIVO2 — Added a direct (feature-less) LiDAR-Inertial-Visual Odometry pipeline (fast-livo2-demo) that fuses LiDAR-inertial pose estimation with dense visual-inertial tracking for robustness in geometrically- or visually-degraded environments. Ported from ROS 1/catkin to ROS 2 (validated on Humble and Jazzy), with a Livox Mid-360 + Intel® RealSense™ D415 sensor profile and validation via public NTU VIRAL dataset replay.

  • LIO SLAM: Point-LIO — Added a high-bandwidth, robust LiDAR-Inertial Odometry pipeline (point-lio-demo) using a point-by-point EKF update (no in-frame motion distortion) and a stochastic-process-augmented kinematic model that tolerates IMU saturation during aggressive motion. Ported to ROS 2/ament with new Avia, Mid-360, and Velodyne configurations and validation via public UrbanLoco dataset replay.

  • GR00T N1.7 (OpenVINO Toolkit) — Added gr00t-n1d7-ov, an implementation of NVIDIA Isaac-GR00T N1.7 embodied AI foundation model for robot manipulation and generalist control, with an end-to-end pipeline that accelerates inference on Intel platforms via OpenVINO IR model conversion. Includes a LIBERO finetuned-checkpoint workflow for convenient evaluation.

  • GEAR-SONIC Whole-Body Control — Added gr00t-wbc, a comprehensive optimization of the SONIC humanoid whole-body-control (WBC) inference pipeline on the Intel® Core™ Ultra “Panther Lake” platform. SONIC is a behavior foundation model that produces natural, whole-body movement (walking, crawling, teleoperation, multi-modal control) from a single unified policy. The pipeline adds OpenVINO inference acceleration, a real-time control thread design, and priority-based NPU scheduling, demonstrating that PTL can meet SONIC WBC’s determinism requirements with substantial power savings compared to GPU execution (OpenVINO 2026.3, ROS 2 Jazzy, Ubuntu 24.04 RT).

Improved

  • Model Predictive Control (MPC) demo — Updated the OCS2-based MPC pipeline (ACT reference model + OCS2 optimization + MuJoCo simulation) documentation and setup for the 2026.2 platform.

  • Added reference Intel® Core™ Ultra “Panther Lake” core-pinning and CPU frequency-locking scripts and reproducible one-command validation flows (reproduce_all.sh) across the new SLAM pipelines for deterministic, comparable benchmarking.

  • Refreshed the imitation-learning and VLA sample pipelines (ACT, Diffusion Policy, Improved 3D Diffusion Policy, Pi0.5 with Real-Time Chunking, Robotics Diffusion Transformer, LLM Robotics, OpenClaw + AgenticROS, ORB-SLAM3) for the 2026.2 documentation and platform references.