← Paper Digest

Multi-CONDOR: Learning Multistable Motion Primitives from Demonstration

Gabriele Giuseppe Di Marzo, Mariano Ramírez Montero, Maximilian Stölzle, Rodrigo Pérez-Dattari, Cosimo Della Santina

November 2026

TL;DR Learns one stable motion policy from demonstrations that converges to any of several goals or to a continuous goal contour. On 2D benchmarks no start state ends away from a goal (StableNODE leaves 0.31% and 7.29%), and trajectory error is about 72% lower than per-goal models picked by a state machine.

What they do

Motion primitives built as stable dynamical systems usually converge to a single goal. Many tasks have several valid endpoints (either handle of a container) or a continuum of them (any point along a rim). Multi-CONDOR extends CONDOR to learn one autonomous vector field from demonstrations that converges either to several discrete goals or to a closed goal contour. The current state alone decides which goal it reaches, with no phase variable or state machine.

How it works

  • A network encodes task states into a 300-dimensional latent space and decodes latent states into task velocities. A multi-step imitation loss reproduces demonstration segments.
  • A stability loss makes encoded rollouts follow prescribed latent reference dynamics. A separation loss stops consecutive states from collapsing onto the same latent point.
  • There are two reference systems. Discrete goals use gradient ascent on a sum of Gaussians, with goal encodings anchored to fixed, well-separated centers. Contours use a radially attracting unit circle, with added curve and orientation losses.
  • A Lyapunov argument shows that trajectories converge to the goal set if the encoded dynamics match the reference exactly.

Results

  • The benchmark is seven hand-drawn 2D tasks with two or three goals, run over 5 seeds. With multi-CONDOR, no start state on the grid ends more than 5 units from every goal. StableNODE leaves 0.31% / 7.29% of start states there, and per-goal CONDORs with a state machine leave 9.22% / 13.75%.
  • Against that state-machine baseline, RMSE falls from 77.85 to 21.58 (two goals) and from 49.70 to 14.08 (three goals). StableNODE’s imitation error is similar or lower (17.65 and 13.46). An imitation-only ablation fits the demonstrations slightly better but leaves 17% and 37% spurious attractors.
  • A Franka FR3 learned from ten kinesthetic demonstrations. It grasps a box by either handle depending on where it starts, approaching from the side as demonstrated.
  • In qualitative 2D examples, the policy converges to star- and circle-shaped goal contours.

Limitations

The convergence guarantee needs exact dynamics matching, and training only encourages it. Closely spaced Gaussians can shift or merge attractors. Discrete goals and contours use separate formulations. The robot experiment shows example executions without a success rate, and contours are shown only in 2D.

Figures

From Figure 1: An encoder maps task states to latent states and a decoder maps them to task velocities. Training alig...
From Figure 1: An encoder maps task states to latent states and a decoder maps them to task velocities. Training aligns the encoded dynamics with latent reference dynamics that have several attractors, while reproducing the demonstrations.
From Figure 3: The latent reference dynamics in two dimensions: two isolated stable attractors (left) and a continuou...
From Figure 3: The latent reference dynamics in two dimensions: two isolated stable attractors (left) and a continuous attracting circle (right).
From Figure 4: Learned dynamics of multi-CONDOR (left) and StableNODE (right) on two three-attractor tasks, with demo...
From Figure 4: Learned dynamics of multi-CONDOR (left) and StableNODE (right) on two three-attractor tasks, with demonstrations in white, rollouts in red and attractors in blue. StableNODE adds spurious attractors away from the goals.
From Figure 6: The Franka arm grasping the container by one of its two handles, driven by a single vector field learn...
From Figure 6: The Franka arm grasping the container by one of its two handles, driven by a single vector field learned with multi-CONDOR.
Abstract

Dynamical-system-based motion policies provide a framework for learning reactive robot motions with stability and convergence properties. Many manipulation tasks, however, admit several distinct goals or a continuous set of acceptable outcomes, requiring motion primitives that preserve these alternatives. Accommodating both goal structures within a unified framework with continuous, autonomous dynamics remains an open challenge. We propose a CONDOR-based framework, multi-CONDOR, for learning motion primitives with either multiple discrete attractors or a continuous attracting goal contour. The approach combines imitation learning with losses that encourage agreement between the learned task dynamics and latent reference systems encoding the desired attraction structure. We provide a Lyapunov-based analysis of the convergence properties of the learned motion dynamics. We evaluate the framework on simulated benchmarks with discrete goals, qualitative examples with continuous goal contours, and robotic grasping experiments. In simulation, multi-CONDOR reduces trajectory RMSE by approximately 72% compared with separately learned single-goal motion primitives selected by a state machine. For tasks with two and three goals, respectively, it reduces the fraction of evaluated initial states ending outside the goal tolerance from StableNODE's 0.31% and 7.29% to 0% in both settings. Experiments with a Franka Emika robot arm demonstrate successful grasping of a container by either of its two handles using motions learned from demonstrations. Together, these results show multi-CONDOR combines motion reproduction with the flexibility to converge to alternative valid outcomes.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.