PaperlayerView exact revision on arXiv

Source-linked AI summary

NeurRAFT: Robot Motion Planning via Anchor-Level Flow Matching with Clearance-Aware Preference Tuning

Sibo Tian, Chang Liu, Minghui Zheng, Xiao Liang

arXiv:2608.24026v1cs.RO

TL;DR

Collision-free planning from partial observations remains difficult because imitation learning does not explicitly penalize near-collision trajectories. NeurRAFT uses compact anchor-level flow matching with Jacobian-weighted training and clearance-based preference tuning, improving benchmark performance and transferring zero-shot to a Franka robot. Limitations include static-environment planning and a remaining simulation-to-real gap.

  • Problem

    Imitation learning from successful demonstrations does not explicitly penalize trajectories that graze or penetrate obstacles, allowing collision-prone behaviors to persist.

  • Method

    NeurRAFT generates compact trajectory anchors with flow matching, uses Jacobian-weighted loss, and directly preference-tunes the planner toward higher-clearance trajectories.

  • Results

    NeurRAFT outperforms classical and state-of-the-art neural planners across all three MπNets test splits and transfers zero-shot to a physical Franka robot on more than 500 planning problems.

  • Takeaways & Limitations

    Clearance-aware preference tuning improves every evaluated setting on top of pretraining and promotes feasible trajectories without additional inference-time computation.

  • Takeaways & Limitations

    NeurRAFT assumes a static environment and must replan from scratch when the environment changes during execution; simulation-to-real robustness also remains limited by perception noise and unseen geometries.

Abstract

from arXiv · show

Recent end-to-end neural motion planners generate trajectories from raw sensor observations, avoiding the privileged geometric models required by classical planners. However, collision-free planning in cluttered environments remains challenging. We present NeurRAFT, a generative planning framework based on anchor-level flow matching and clearance-aware preference tuning. Unlike prior neural planners that model dense waypoint sequences and spend capacity on redundant local details and smoothness, NeurRAFT operates on compact anchor waypoints. We train the planner using a Jacobian-weighted loss that accounts for the task-space impact of each anchor. At inference, the anchors are generated in two integration steps, followed by cubic-spline interpolation to recover a smooth, full-resolution trajectory. Since imitation learning from positive demonstrations cannot distinguish collision-free from near-collision trajectories, collision-prone behaviors persist at test time. Rather than relying on post-hoc corrections, we directly reshape the pretrained planner's distribution toward safer solutions without augmenting inference. Specifically, Direct Preference Optimization shifts probability mass toward trajectories with larger obstacle clearance, with the resulting improvement directly absorbed into the planner parameters. Experiments show substantial improvements over state-of-the-art planners, while real-world experiments demonstrate zero-shot transfer to a Franka robot under noisy and partially occluded depth observations. Video results available at https://neurraft.github.io/.

I. INTRODUCTION

NeurRAFT addresses cluttered, partially observed robot motion planning with compact anchor-level flow matching and clearance-aware preference tuning. It combines task-space-aware training, two-step anchor generation, spline reconstruction, and distribution-level safety alignment.

  • Representation and generation: A Jacobian-weighted flow matching loss accounts for how anchor errors affect end-effector displacement.The weighting addresses kinematic differences across joints and configurations.
  • Representation and generation: NeurRAFT generates compact on-trajectory anchors with flow matching instead of modeling dense waypoint sequences.The full-resolution trajectory is recovered by cubic-spline interpolation.
  • Representation and generation: Inference generates anchors in only two integration steps and reconstructs a smooth trajectory through cubic-spline interpolation.This avoids modeling redundant local smoothness directly.
  • Preference alignment: Clearance-aware preference alignment reshapes the planner’s trajectory distribution toward safer solutions rather than applying post-hoc inference-time corrections.The planner’s own samples are ranked by obstacle clearance to provide negative supervision absent from imitation learning.
  • Evaluation: Experiments on MπNets show substantial planning-success improvements, with zero-shot transfer to a Franka robot under noisy and partially occluded depth observations.The real-world result is reported alongside benchmark comparisons against state-of-the-art planners.

II. RELATED WORK

Related neural planners either guide classical search, predict trajectories autoregressively, or generate complete trajectories with diffusion or flow matching. NeurRAFT differs by using offline, clearance-based preference alignment without online interaction, human involvement, or added inference-time computation.

  • Neural planning paradigms: Learned sampling methods accelerate classical search but retain explicit collision checking in the planning loop.They learn sampling distributions rather than directly replacing the search structure.
  • Neural planning paradigms: Autoregressive planners predict successive configurations directly, but multi-step rollout can accumulate prediction errors over the planning horizon.This contrasts with trajectory-level generative approaches that produce complete trajectories at once.
  • Generative planners: Diffusion planners model multimodal full trajectories but typically require many iterative denoising steps, whereas flow matching enables fewer integrations.Conditional generation preserves diverse feasible motion modes.
  • Inference-time correction: State-of-the-art methods commonly use post-hoc refinement, guidance, hierarchical planning, or trajectory optimization to improve planning success.These mechanisms operate after or during inference rather than directly reshaping the learned output distribution.
  • Preference alignment: NeurRAFT uses signed obstacle clearance from its own samples to construct offline preference pairs, requiring neither online interaction nor human involvement.The approach targets safer trajectories without introducing additional inference-time computation.

III. METHODOLOGY

NeurRAFT formulates motion planning as conditional generation over compact anchor sequences derived from point-cloud observations and boundary configurations. A deterministic cubic spline converts anchors into smooth trajectories while preserving the start and goal constraints.

  • Problem formulation: The planner receives a scene point cloud, start configuration, and goal end-effector pose, then produces a collision-free trajectory without a privileged geometric model at inference.Obstacle information must be extracted from the raw point cloud.
  • Anchor representation: The method replaces a dense H-waypoint trajectory with N anchor waypoints, where N ≪ H.This reduces the generation dimensionality from H × d to N × d.
  • Anchor representation: The first and last anchors are fixed to the start and goal configurations, while each intermediate anchor lies on the executed trajectory.The full trajectory is recovered by fitting a cubic spline through the anchors.
  • Trajectory reconstruction: The spline reconstruction yields a twice-continuously differentiable trajectory with exact boundary conditions, which is then discretized into H uniformly sampled waypoints.Smoothness and endpoint satisfaction follow from the reconstruction design.
  • Generative formulation: Conditional generation preserves multimodal feasible trajectories, allowing diverse candidates for scenes with topologically distinct solutions.A deterministic predictor could average across modes and produce an invalid intermediate path.

B. Anchor-Level Flow Matching

Anchor-level flow matching learns a conditional distribution over compact anchor sequences by connecting Gaussian source samples to expert anchors along linear probability paths. Kinematics-aware weighting and fixed endpoints tailor training and sampling to robot execution constraints.

  • Flow matching: Expert trajectories are subsampled at uniform arc-length spacing into N anchor configurations, while start and goal anchors remain fixed.Flow matching is trained on these compact anchor-sequence samples.
  • Flow matching: The probability path linearly interpolates Gaussian source samples A(0) and expert anchor sequences A(1), with straight-line displacement as the target velocity.All intermediate anchors share a flow time during training and sampling.
  • Boundary handling: Endpoint anchors are clamped to qs and qg after every integration step, enforcing the boundary conditions exactly.The endpoint anchors carry zero target velocity during training.
  • Network parameterization: The model predicts the clean anchor sequence from a partially noised sequence, context, and per-anchor flow times, from which velocity estimates are recovered.The implementation trains the recovered velocity against the flow-matching target.
  • Kinematics-aware loss: The kinematics-aware term weights anchor errors by their task-space effect, complementing the standard flow matching loss.The final objective combines the flow matching and kinematics-aware terms.
  • Inference: At inference, an explicit Euler scheme integrates the learned flow, while multiple goal configurations generate parallel candidate trajectories for redundant manipulators.Candidates are collision-checked against a mesh reconstructed from the observed point cloud.

C. Clearance-Aware Preference Tuning

NeurRAFT uses signed-clearance preferences to distinguish safer trajectories from collision-prone behaviors that positive-only imitation learning cannot penalize. DPO then reshapes the planner toward higher-clearance solutions while retaining expert behavior and leaving inference unchanged.

  • Positive-only imitation learning does not penalize trajectories that graze or penetrate obstacles, so collision-prone behaviors can persist at test time.
  • Candidate anchor sequences are reconstructed into dense trajectories and scored using the minimum signed clearance across the planning horizon.Positive signed distances indicate collision-free clearance, while negative values indicate penetration.
  • Each context forms a preference pair whose winner has maximum minimum clearance and whose loser has minimum minimum clearance.This separates feasible from infeasible trajectories when only some candidates are collision-free.
  • DPO uses a flow-matching surrogate for trajectory log-likelihood because exact likelihoods are intractable for the flow-based planner.The preference loss is evaluated on interior anchors because pinned endpoints are identical across candidates.
  • The combined objective preserves expert behavior while reshaping the learned distribution toward trajectories with larger geometric clearance, without changing sampling or adding inference-time refinement.Privileged obstacle geometry used to construct preferences is restricted to training.

A. Experimental Setup

NeurRAFT is evaluated on the MπNets benchmark for a 7-DoF Franka manipulator across three cluttered-scene test splits. The study compares classical and neural planners under multiple demonstration sources and candidate-generation regimes.

  • The MπNets benchmark evaluates a 7-DoF Franka manipulator in tabletop, cubby, and dresser scenes.Demonstrations come from global AIT* planning or a hybrid AIT* plus Geometric Fabrics pipeline.
  • A trial succeeds when the dense reconstructed trajectory reaches the target pose while remaining collision-free and satisfying joint-limit and self-collision constraints.Success is verified in simulation using ground-truth scene geometry and a high-fidelity collision model.
  • The evaluation varies both the number of sampled trajectories and the number of feasible inverse-kinematics goals.The four regimes range from one sample for one goal to five samples for every feasible goal.
  • Table I reports success rates on the Global-Solvable, Hybrid-Solvable, and Both-Solvable test splits.
  • Classical baselines include global and hybrid planners, CHOMP, Geometric Fabrics, and STORM, alongside recent neural motion planners.
  • The default implementation samples five trajectories from each feasible inverse-kinematics goal.Training and benchmark evaluation use an NVIDIA A100 GPU; inference and real-world experiments use an RTX 4080.

B. Comparison with Baselines

NeurRAFT achieves high success rates across all MπNets splits and consistently outperforms classical and neural baselines. Preference tuning further improves every evaluated setting, while qualitative results show larger-clearance trajectories without post-hoc refinement.

  • 93.61%, 97.89%, and 97.94% are NeurRAFT’s success rates on the three splits with Global Expert training.With Hybrid Expert training, the corresponding rates are 89.56%, 98.22%, and 98.17%.
  • NeurRAFT exceeds 89% on every split, while classical planners range from 26.67%–50.22%, 31.61%–78.44%, and 32.20%–76.00% across the three splits.Unlike NeurRAFT, classical planners query privileged geometric models and perform collision checking during planning.
  • On the Global-Solvable split, NeurRAFT exceeds Cascaded Diffusion by 8.48 percentage points with Global Expert training and Neural MP by 13.23 points with Hybrid Expert training.This split is the most challenging setting for every planner.
  • Across three representative cases, preference alignment shifts trajectories toward higher clearance and produces safer paths without post-hoc refinement.
  • The results support compact anchors and clearance-based preference tuning as the two-stage NeurRAFT design.The pretrained planner already matches or exceeds published baselines, while preference tuning improves performance across settings.
  • Preference tuning improves success in all six settings, adding 4.05–4.39 points on Global-Solvable, 4.72–1.39 points on Hybrid-Solvable, and 3.61–1.50 points on Both-Solvable.The two settings where pretraining alone fell below the strongest baseline surpassed it after tuning, with no degradation.

C. Effect of Candidate Diversity

Increasing candidate diversity improves NeurRAFT’s success rate across all evaluated sampling regimes. Diversity is varied through both the number of feasible goals and the number of samples per goal.

  • Enlarging the candidate set improves success rate in every candidate-generation setting.The study varies samples per goal and the number of inverse-kinematics goal configurations.

D. Ablation Studies

Ablations show that compact anchors preserve planning quality while reducing inference cost, and Jacobian weighting plus preference tuning improve the learned trajectory distribution. Increasing preference strength improves both success and feasibility, especially when candidate diversity is limited.

  • Anchor budget: Across N = 4–50 anchors, success and feasibility remain within a few percentage points except at N = 4, while computational cost grows linearly with N.The study adopts N = 8 as a cost-effective setting.
  • Jacobian-weighted loss: The Jacobian-weighted loss improves feasibility in 23 of 24 cases and success in 20 of 24 cases, with gains up to 2.32 and 2.28 percentage points, respectively.In the remaining comparisons, performance differences are at most 0.66 percentage points.
  • Candidate diversity: Enlarging the candidate set improves success in every setting, with larger gains when candidates use multiple IK solutions rather than repeated samples.The four settings vary feasible IK solutions and trajectories sampled per goal.
  • Preference strength: At β = 0, continued standard imitation training changes performance by only 0.02 percentage points on average, with no consistent direction.The largest deviation from the pretrained planner is 1.72 points.
  • Preference strength: As β increases from 0 to 100, all success and feasibility metrics improve and gradually converge across expert settings and candidate-generation configurations.For the Global Expert on the Global-Solvable split, single-sample gains reach 14.44 percentage points for both metrics.
  • Preference strength: With multiple IK solutions, feasibility gains exceed success gains because clearance tuning improves individual samples while success saturates when candidate sets already contain a solution.For the Global Expert, feasibility gains are 18.88 and 18.30 points, versus success gains of 7.50 and 4.17 points.

E. Real-World Experiments

Real-world experiments evaluate simulation-trained NeurRAFT on Franka scenes of increasing complexity without retraining or adaptation. Preference tuning improves every task, and the planner transfers zero-shot despite noisy and partially occluded depth observations.

  • Experimental setting: The physical evaluation uses tabletop, drawer, and shelf scenes with varied obstacles, placements, start configurations, and target end-effector poses.The deployment study contains more than 500 real-world planning problems.
  • Results: Preference tuning improves performance on every real-world task, with average gains of 6.39 points for the Global Expert and 12.93 points for the Hybrid Expert.The same improvement trend appears in the benchmark results.
  • Results: NeurRAFT achieves average success rates of 86.04% and 86.75% for the Global and Hybrid Expert settings, respectively, under noisy and partially occluded depth observations.The reported performance is approximately 10 percentage points lower than in simulation.

V. CONCLUSION

NeurRAFT combines compact anchor-waypoint generation, Jacobian-weighted training, spline recovery, and clearance-aware preference tuning to produce safer motion plans. It outperforms classical and neural baselines across benchmark settings and transfers zero-shot to a physical Franka manipulator, while remaining limited to static environments and affected by the simulation-to-real gap.

  • NeurRAFT generates compact anchor waypoints with flow matching, fixes the start and goal anchors, and recovers smooth trajectories through cubic-spline interpolation.Its training objective includes a Jacobian-weighted term that scales joint-space errors by their induced end-effector displacement.
  • Clearance-aware preference tuning ranks the planner’s own samples by signed obstacle clearance and directly fine-tunes the model toward safer paths without additional demonstrations or labels.The preference data come from the existing training set and require no online interaction.
  • NeurRAFT outperforms classical and state-of-the-art neural planners on all three MπNets test splits under both expert data pipelines.The pretrained planner already exceeds the strongest published baselines in most settings, and preference tuning improves every setting further.
  • More than 500 real-world planning problems confirm zero-shot transfer to a physical Franka manipulator under noisy and partially occluded depth observations.The ordering between the pretrained and preference-tuned stages is preserved in these experiments.
  • NeurRAFT assumes static environments and fixed observations, requiring replanning from scratch when conditions change; simulation-to-real differences also constrain robustness to noise and unseen geometries.The paper identifies reactive planning, domain randomization, and real-world preference data as possible future directions.
Loading 2608.24026v1…