ELMP Efficient Learning for Motion Planning via Analytical Policy Gradients

Jul 16, 2026

Upload by Logic

Neural Motion Planners enable fast reactive motion, but adapting to new environments often needs recollecting large expert datasets, making it computationally prohibitive. ELMP proposes data-efficient adaptation through self-supervised fine-tuning: it directly optimizes a neural policy via a differentiable kinematic layer with dense collision, target-reaching, and smoothness objectives. Tool geometry is explicitly encoded with point clouds to generalize across changing kinematic chains. Benchmarks show 84.8% average success and large reductions in cold-start latency, and unseen environments improve success from 57.3% (zero-shot) to 89.8%, validated on a Franka Emika Panda robot.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

Introduction

Related Work

ELMP: Efficient Learning for Motion Planning via Analytical Policy Gradients Yixiao Li 1 , Tifanny Portela 1 , Jordis Herrmann2 ,†, Ren Zurbr¨ugg 1 , Marco Hutter1 arXiv :2607 .00215v1 [ cs .RO] 30 Jun 2026 Abstract—Neural Motion Planners (NMPs) enable fast reactive motion generation, but adapting them to new environments typically requires recollecting large expert datasets, which is computationally prohibitive. We propose ELMP, a framework for data-efficient adaptation via self-supervised fine-tuning. Rather than generating additional expert trajectories with expensive global planners, ELMP directly optimizes the policy through a differentiable kinematic layer using dense collision, target-reaching, and smoothness objectives. This replaces expert data generation with rapid problem sampling, reducing persample adaptation cost by roughly two orders of magnitude. To further support robust generalization across changing kinematic chains, we introduce a mechanism to explicitly encode tool geometry via point clouds. Benchmarked against classical and neural baselines, ELMP achieves an 84.8% average success rate with orders-of-magnitude lower cold-start latency than classical methods. In unseen environments, self-supervised fine-tuning improves success rate from 57.3% (zero-shot) to 89.8%, removing the data collection bottleneck. Our approach maintains millisecond-level inference latency and is validatedon a physical Franka Emika Panda robot. Index Terms—Robot Learning, Motion Planning, Collision Avoidance, Analytical Policy Gradient I. INTRODUCTION Generating collision-free motion is fundamental to robotic manipulation but remains challenging in unstructured environments [1]–[3] . Traditional sampling-based [4]–[8] and optimization-based [9]–[12] methods provide completeness or optimality guarantees but incur high latency from their sequential sense-plan-act pipeline, often precluding real-time replanning. Local reactive methods such as STORM [13] and Geometric Fabrics [14] achieve high-frequency control but lack long-horizon foresight, making them vulnerable to local minima in complex geometries. Neural Motion Planners (NMPs) [1]–[3] compress planning into a neural network for fast inference, but standard Behavior Cloning (BC) suffers from covariate shift over long horizons, requiring prohibitive amounts of expert data. Moreover, most NMPs assume fixed robot geometry, neglecting variable end-effectors or grasped objects. Although recent methods such as Neural MP [2] have begun addressing variable embodiments, they lack extensive evaluation of pointcloud representations, and bounding-box abstractions [15] sacrifice geometric fidelity, failing in cluttered spaces. For tool-aware manipulation (Fig. 1), explicit geometric reasoning about the full kinematic chain is required. This work was primarily supported by the ETH AI Center. 1Robotic Systems Lab, ETH Z¨urich; 2ABB. †Jordis Herrmann is currently with ANYbotics. Email: {yixili,[tportela](tportela}@ethz.ch)[}](tportela}@ethz.ch)[@ethz.ch](tportela}@ethz.ch) Fig. 1: Tool-Aware Manipulator Motion Planning: AFranka Emika Panda robot executes a pick-and-place task involving a tool (wrench) . Our method ELMP explicitly encodes the variable tool geometry via point clouds to enable collision avoidance for the entire kinematic chain, while leveraging Analytical Policy Gradients to fine-tune the policy for high-precision, collision-free motion. We introduce ELMP (Efficient Learning for Motion Planning via Analytical Policy Gradients), a two-stage neural planner: BC pre-training followed by self-supervised APG fine-tuning through a differentiable kinematic layer with dense collision, target-reaching, and smoothness objectives. This enables adaptation to novel environments without additional expert demonstrations. We further encode tool geometry as a point cloud, conditioning the policy to generate collision-free trajectories for the entire kinematic chain. While ELMP bui

FAQs

What problem does ELMP address in neural motion planning?

ELMP addresses the high cost of adapting neural motion planners to new environments, which typically requires recollecting large expert datasets and causes computational bottlenecks and cold-start latency.

How does ELMP fine-tune the policy without collecting additional expert trajectories?

ELMP uses self-supervised fine-tuning with Analytical Policy Gradient (APG) through a differentiable kinematic layer, optimizing dense collision avoidance, target-reaching, and smoothness objectives while replacing expensive expert data generation with fast problem sampling.

How does ELMP handle variable tool geometry during motion planning?

ELMP explicitly encodes tool geometry using point clouds, conditioning the planner to reason over the full kinematic chain and improve collision avoidance in cluttered environments.

Jul 16, 2026

Upload by Logic