Trajectory Planning without Trajectory Data: A Manifold-Guided Approach

AI summary

Overview

Research area: Robotics — learning-based motion planning and trajectory generation, connecting generative modeling (diffusion / score-based models) with the geometry of learned data manifolds.

Technical level: Intermediate. The core intuition (walk along the data manifold instead of imitating trajectories) is approachable, but the paper's central construction is an ordinary differential equation derived through operator splitting, plus Hessian-based normal-space estimation, which requires comfort with calculus, differential geometry vocabulary, and ODE integration.

Scope: The paper proposes Ariadne, a planner that trains only on valid states (no trajectory demonstrations) and builds paths at inference time by exploiting the geometry of the state-space manifold learned by a generative model, evaluated on Maze2D, a 6-DoF tabletop arm (YAM 6D), and dual 7-DoF arm planning (DualKUKA).

What This Paper Is About

Most learning-based planners are trained on large collections of expert trajectories and then generate new trajectories by conditioning on a start and goal. Because the space of possible trajectories is combinatorially large, finite datasets cover it poorly, and models generalize badly to start–goal pairs whose connecting path was never observed — even when both endpoints were seen during training. Ariadne asks whether planning is possible with state observations alone, by learning the low-dimensional manifold that feasible states lie on and then constructing a path by following that manifold's geometry from one endpoint to the other.

Key Contributions

1. Extends the manifold hypothesis for score-accessible generative models. Rather than assuming a simple linear subspace or a (D−1)-dimensional manifold, the paper extracts general local geometry (tangent and normal spaces) from a trained diffusion/score model using eigendecomposition of the Jacobian of the score function — without strong structural assumptions about the data manifold.

2. A trajectory planning framework requiring only state-space training data. Ariadne introduces an endpoint constraint by connecting a non-linear interpolation formulation with an iterative projection method, so that paths are built at inference time rather than imitated from data.

3. The Correction-Projection ODE (CP-ODE). The authors reinterpret the naive "move-then-project" iteration as the discretization of an ODE via operator splitting, then analytically add a correction term so that the integral curve both stays on the manifold and satisfies the prescribed start and end points.

4. Evaluation against three families of baselines. Ariadne is compared against classical search (RRT), trajectory-supervised diffusion planners (Diffuser), geometry-based trajectory optimization (Follow-the-Energy), and a classical planner (PBDMP) across Maze2D and robotic motion-planning tasks, showing competitive performance from state-only supervision.

Main Findings

• State-only training can beat trajectory-supervised planning on unseen start–goal pairs. On Maze2D, Ariadne scores 55.69 (Medium) and 50.63 (Large) in D4RL normalized score, versus 52.70 and 41.98 for Diffuser* — the variant whose training data excludes trajectories reaching the evaluation target region. Full-supervision Diffuser reaches 120.52 and 120.37, which the authors present as evidence that directly observing solutions to the evaluation tasks is a large advantage.

• Manifold construction works beyond 2D mazes. On the YAM 6D tabletop manipulation task, Ariadne raises collision-free planning success from 12.31% (straight-line reference) to 60.77%.

• Competitive on high-dimensional dual-arm planning. On DualKUKA, Ariadne achieves the highest average success rate (53.3%) across the three environments, compared with PBDMP* 53.2%, RRT 44.7%, Follow-the-Energy 42.3%, Diffuser 31.7%, and Straight Line 21.3%. Per-environment: 57.5 (n=7, s=0.14), 50.5 (n=5, s=0.14), 52.0 (n=5, s=0.22).

• Local correction beats trajectory-level geometric optimization in narrow-feasible-region environments. Follow-the-Energy scores 16.29 and 16.41 on Maze2D Medium and Large, and 30.0, 44.0, 53.0 on the three DualKUKA settings. The authors attribute this to trajectory-level interpolation producing overly smooth paths that leave narrow feasible regions, whereas Ariadne applies repeated local manifold-guided corrections.

• The two ODE terms play complementary roles, verified on a known ground-truth manifold. On the UCG (uniform circular Gaussian) example where the true semicircular manifold is known, Straight Line has manifold distance 0.8800 and endpoint error 0.0000; Projection-Only has manifold distance 0.0322 but endpoint error 3.3652; Ariadne has manifold distance 0.0798 and endpoint error 0.3998 — substantially reducing manifold deviation versus the straight line while reducing endpoint error by an order of magnitude versus Projection-Only.

• Trajectory supervision quality matters in the dual-arm domain. The paper notes that Diffuser's expert trajectories are produced by a classical planner and need not be short or optimal, whereas Ariadne does not imitate a trajectory distribution and instead constructs paths from the distribution of feasible states.

• Robustness to retraining and to hyperparameters. The authors evaluate sensitivity to score-model retraining across five independently trained models and report consistent performance. Ablations on DualKUKA (n=7, s=0.14) show the method does not rely on narrow tuning: normal threshold −12/−20/−25/−30/−35 gives 49.5/57.5/53.5/51.5/49.0; step size 0.01/0.02/0.03/0.04 gives 54.5/57.5/57.5/54.0; projection scale 0.4/0.5/0.6/0.7/0.9 gives 52.0/57.5/54.5/53.0/50.0.

Methodology in Plain English

The paper starts from an observation: even when a dataset of trajectories is incomplete, the states those trajectories pass through tend to be covered much more broadly. So instead of learning "what trajectories look like," Ariadne learns "what valid states look like," using a generative model trained on state observations only.

Feasible states are assumed to cluster on a lower-dimensional, possibly curved surface (the data manifold) sitting inside the high-dimensional state space. The plan is to build a path that stays on that surface and still connects a given start and goal.

Three pieces make this work:

1. Finding which way is "off the surface." The score function of the generative model — the gradient of the log density — points toward high-density regions and works as a normal-direction approximation when the manifold has co-dimension one. For general, lower-dimensional manifolds, the authors instead use the Hessian (second derivative) of the log density: directions of large negative curvature are normal to the manifold, and near-zero curvature directions are tangent.

2. A notion of distance. They derive a Riemannian metric from the score, G(x) = exp(‖s(x)‖)/Z · I, which stays near the identity on the manifold and grows away from it, penalizing off-manifold deviations.

3. A path that both hugs the surface and lands on target. The naive procedure — step toward the goal, then project the displacement back onto the tangent space — keeps the path near the manifold but shrinks progress toward the goal, sometimes stalling. The authors show this iteration is a discretization of an ODE (viewing "move" and "project" as two split operators A and B), then analytically add a correction term, (2−4t)·φ(x₀,x₁,t), that gradually adds back the removed components. The resulting Correction-Projection ODE is solved numerically with Euler steps and operator splitting.

At inference, the planner simply integrates this ODE from the start state to the goal state, producing a state sequence. Practical choices include using a straight-line reference path between endpoints and arc-length-based step sizes. Three hyperparameters matter — the initial arc-length/step size, the projection scale ε, and the normal selection threshold (how negative an eigenvalue must be to count as a normal direction). The authors state that discretization and imperfect geometry estimation can introduce small residual manifold drift and endpoint error, which Table 2 quantifies.

Why This Matters

Impact on research. The paper reframes trajectory planning as manifold-constrained curve construction rather than distribution matching over trajectory sequences. If planning can be learned from state coverage instead of demonstrations, the data-collection bottleneck for robot learning shifts: state data is generally cheaper and less restrictive to gather than full expert trajectories. It also connects generative-model geometry (score, Hessian, learned metrics) directly to a control/planning objective.

Real-world applications:

• Dual-arm industrial manipulation, where two 7-DoF arms must avoid both workspace obstacles and each other — the setting where the paper reports the highest average success rate.
• Tabletop pick-and-place and assembly, represented by the YAM 6D collision-free manipulation benchmark.
• Navigation in cluttered indoor or warehouse environments, analogous to the Maze2D setting with start–goal pairs not seen during training.
• Any deployment where expert demonstrations are scarce or biased but system telemetry or safe-state logs are plentiful — the state-only supervision regime.

Industry relevance. Training data cost and generalization to novel start–goal queries are two persistent pain points in deployed motion planning. A method that learns from readily available state logs, produces no trajectory imitation bias, and remains competitive with classical search (RRT) and scene-aware planners (PBDMP) is relevant to robotics companies doing manipulation, logistics, and autonomous navigation. The paper's own caveat is important for practitioners: Ariadne learns one planner per scene, whereas PBDMP takes scene layout as input to generalize across environments.

Future Directions

• Extending beyond state validity to transition validity. The authors flag that state validity does not guarantee transition validity in systems with irreversible, discontinuous, or strongly action-dependent dynamics, and identify extending the model to state-action or transition tuples as an important direction.

• Better reference paths for long-horizon and multimodal problems. Planning quality can degrade when state coverage is insufficient, when connectivity is multimodal, or in long-horizon problems where a simple reference path fails to capture the correct global topology. More informative reference paths are called out as future work.

• Scaling manifold-geometry estimation. Because estimating the normal space requires derivatives of the learned score field, inference becomes more computationally expensive as state dimension grows.

• Cross-environment generalization. The comparison with PBDMP highlights that PBDMP uses scene information while Ariadne learns a separate planner per scene; generalizing across layouts and extending the framework to more complex environments remain open.

Target Audience

Robotics and machine-learning researchers working on motion planning, diffusion/score-based generative models, and manifold learning; graduate students interested in geometry-informed control; and applied engineers evaluating whether state-only supervision can replace trajectory demonstration collection. Readers wanting the full derivations, dataset construction details, evaluation protocols, efficiency comparisons, and the retraining study should consult the appendices referenced in the paper (Appendices B, D, E, F, I, K, L, M).

Authors’ abstract

A common way for trajectory planning is to leverage generative models trained on large collections of expert trajectories. At inference time, the model generates executable trajectories by conditioning on task goal constraints. However, trajectory-based methods rely on costly supervision, scale poorly with sequence length, and often generalize poorly to unseen constraints such as novel start-goal pairs. We propose an alternative to learn the underlying state-space manifold and use the geometry of the manifold for trajectory planning. This approach requires only state observations and enables generalization to unseen constraints by con- structing trajectories on the learned manifold of the state space. Experiments on Maze2D and robotic motion-planning benchmarks show that Ariadne constructs feasible paths from state-only supervision and generalizes to unseen start-goal combinations. On high-dimensional dual-arm planning, it remains competitive with trajectory-supervised and classical planners, while requiring no trajectory data for training.