Spline Policy: A Structured Representation for Robot Policies
Mengze Tian, Yiming Li, Sichao Liu, Auke Ijspeert, Sylvain Calinon
cs.RO
2026-06-05
Spline Policy keeps the backbone and swaps action chunks for spline parameters: scores hold across six tasks, network FLOPs drop ~30%, and perturbation recovery improves sharply.
Modern imitation-learning policies for robot manipulation all represent actions as fixed-resolution action chunks: a discrete sequence of waypoints predicted in one shot. ACT, Diffusion Policy, Flow Matching Policy, and VLA models such as π0 are no exception. The format is simple and effective, but it hides structure. Before execution there is no explicit continuity, no derivatives, no boundary conditions, no way to resample at a different temporal resolution, and no path to propagate uncertainty or hook into a classical controller.
The learning-from-demonstration tradition has the opposite strength. Movement primitives such as DMPs and ProMPs carry geometric and temporal structure, are easy to edit and constrain, and connect naturally to control. But they do not fit modern high-dimensional, multimodal, perception-conditioned backbones. The authors (Ijspeert at EPFL and Calinon at Idiap, both long-time movement-primitive researchers) set out to merge the two: keep the backbone, change only the output representation.
The move is one line. The policy backbone (diffusion, flow-matching, transformer, or VLA, your choice) still maps observations to actions, but instead of emitting a chunk of waypoints it emits a set of spline parameters w. Decoded with spline basis functions, this gives a continuous trajectory f(t) = φ(t)·w that you can sample at any temporal resolution.
That spline enables things an action chunk cannot:
The second realization is the one worth explaining. With a quadratic Bernstein spline, an analytical distance-field construction turns the spline into a state-dependent flow field: for the current state x, project to the nearest point on the spline, then add an attraction term toward the spline and a tangential progression term along it. Under regularity assumptions the induced dynamics never increase the distance from the state to the spline, so a pushed robot is pulled back to the planned motion — a principled local correction mechanism. This flow field can also be combined with a classical collision-avoidance controller through null-space projection, with no retraining of the backbone.
On the LASA dataset the authors test perturbation recovery (25 perturbations × 3 demonstration types) with three metrics: Chamfer distance, convergence error, and maximum speed.
| Method | Chamfer (mm) | Conv. error (mm) | Max speed (m/s) |
| Baseline Diffusion Policy | 26.2 | 26.7 | 13.0 |
| SP (Trajectory) | 21.0 | 28.8 | 14.3 |
| SP (Flow) | 12.8 | 1.1 | 0.25 |
The flow-field variant cuts convergence error from 26.7 mm to 1.1 mm. That is the headline number of the paper.
Under injected observation noise, the probabilistic SP keeps Chamfer distance far lower: at 10 mm noise the baseline scores 8.6 and the probabilistic trajectory variant 0.7; at 40 mm noise it is 20.4 versus 6.1.
The behavior-cloning benchmark covers six tasks across state, vision, and point-cloud inputs, with backbones strictly matched. Be honest here: SP is not better on task score. Some entries are flat, some slightly better, some slightly worse (on Tool Hang, SP-Diff's 0.80/0.76 trails the baseline's 0.88/0.80; on Can everyone reaches 1.0). The authors state plainly that this benchmark only shows swapping the output keeps performance in the same range, not that it is superior. The consistent change is compute: because the output shrinks from a full waypoint chunk to 8 spline parameters per action, network forward FLOPs fall to 50.67%–87.63% of baseline depending on the task (state and point-cloud tasks drop most), averaging roughly a one-third cut.
Real-robot case studies on a Franka cover four scenarios: wine-glass delivery trained on 120 human demonstrations, recovery under pushing and pulling, null-space collision avoidance, and gentle shaking of a liquid-filled glass. ALOHA deployments run ACT, Diffusion Policy, and π0.5 backbones on PushT and toy packing with randomized poses.
The value is not a smarter robot; the task scores say it does not get one. The value is a layer of operable motion structure grafted onto a modern policy: lower compute, plus the ability to resample, quantify uncertainty, self-correct locally during execution, and interface with classical control. For an engineer trying to deploy a policy stably on a real robot, these are interface capabilities an action chunk cannot give and that genuinely matter.
The authors are upfront. First, SP does not replace an accurate backbone: if the policy predicts a bad spline, the structured decoder and flow field cannot guarantee task success on their own. Second, the pull-back-to-spline property is local, holding only near the generated motion under the construction's assumptions; it is not a task-level stability or success guarantee and does not cover off-manifold states. Third, SP is less suited to highly discontinuous or dynamic interactions such as hitting a moving object. Fourth, the analytical flow-field path relies on quadratic splines with C⁰/C¹ continuity; extending to broader spline families is left as future work.
One more caveat from the reading: across the six benchmark tasks, cases where SP clearly wins are rare; the main gains concentrate in the low-dimensional LASA perturbation and noise experiments and in the FLOPs reduction. The conclusion holds only if you read this as a representation-level change rather than a new SOTA.