Hdmove2 Patched May 2026

[5] T. P. Lillicrap et al., "Continuous control with deep reinforcement learning," International Conference on Learning Representations (ICLR) , 2016.

[ z^* = \arg\min_z(t) \in Z \mathcalJ[D(z(t))] \quad \texts.t. \quad \texthomotopy constraints ]

[2] N. Ratliff, M. Zucker, J. A. Bagnell, and S. Srinivasa, "CHOMP: Gradient optimization algorithms for efficient motion planning," IEEE International Conference on Robotics and Automation (ICRA) , 2009, pp. 1292–1299. hdmove2

hdmove2 achieves a 94% success rate on the narrow passage benchmark (64 DoF), compared to 12% for RRT* and 68% for CHOMP. The event-driven controller reduces average planning frequency by 63%, enabling 95 Hz control updates.

where ( \mathbfM ) is a configuration-dependent inertia matrix and ( c_obs ) is a smooth barrier function. Instead of solving directly in ( Q ), hdmove2 solves: [ z^* = \arg\min_z(t) \in Z \mathcalJ[D(z(t))] \quad \texts

The lower level is solved using a fast alternating direction method of multipliers (ADMM) that converges in under 5 ms for ( n \leq 128 ). Re-planning is triggered when:

| Algorithm | Success Rate (Bench B) | Planning Time (ms) | Cumulative Jerk (m²/s⁵) | Real-time feasible (>30 Hz) | |-----------|------------------------|--------------------|--------------------------|-------------------------------| | RRT* | 0.12 ± 0.05 | 3420 ± 450 | 18.4 ± 3.2 | No | | CHOMP | 0.68 ± 0.12 | 520 ± 85 | 9.2 ± 1.8 | No (for n>30) | | hdmove1 | 0.71 ± 0.10 | 88 ± 12 | 5.3 ± 0.9 | Yes (at 35 Hz) | | | 0.94 ± 0.04 | 41 ± 6 | 1.4 ± 0.3 | Yes (at 95 Hz) | Zucker, J

[ \mathcalJ[\tau] = \int_0^T \left( \underbrace \dot\tau(t) \textkinetic energy + \lambda_1 \underbrace \dddot\tau(t) \textjerk + \lambda_2 \underbracec_obs(\tau(t))_\textcollision cost \right) dt ]