Research
Active Constraint Learning in High Dimensions from Demonstrations
Overview Research area: Robotics — learning from demonstrations (LfD), inverse optimal control (IOC), constraint inference, Gaussian process (GP) learning, robot safety. Technical level: Advanced. The

- arXiv
- 2512.22757
- Published
- 2025-12-28
- Authors
- Zheng Qiu, Chih-Yuan Chiu, Glen Chou
AI summary
Overview
- Research area: Robotics — learning from demonstrations (LfD), inverse optimal control (IOC), constraint inference, Gaussian process (GP) learning, robot safety.
- Technical level: Advanced. The paper assumes familiarity with Karush-Kuhn-Tucker (KKT) optimality conditions, Gaussian process posteriors and covariance-based uncertainty, constrained trajectory optimization, and inverse optimal control.
- Scope: The paper proposes GP-ACL, an iterative active constraint learning algorithm that queries start and goal states so that a demonstrator produces sparse but maximally informative trajectories, from which an unknown nonlinear constraint is inferred via a Gaussian process.
What This Paper Is About
Robots can learn what an expert is avoiding by watching expert trajectories, but existing constraint-learning methods treat the supplied demonstrations as fixed and generated without regard to the learning task — so the data may be uninformative and a large number of demonstrations may be needed. This paper asks how to actively choose which start and goal states to query so that the resulting demonstrations reveal the most about an unknown constraint. The authors build an iterative loop around a Gaussian process constraint model and show that this actively generated data beats a random-sampling baseline on high-dimensional simulation and hardware tasks.
Key Contributions
- Multiple GP posterior sampling over the constraint. Building on the GP-based constraint learning framework of Chou et al. (2022), the authors sample multiple GP posterior estimates of the unknown constraint given a demonstration dataset D, giving a fuller description of both the constraint information encoded in D and the uncertainty that D cannot resolve.
- The GP-ACL algorithm (Alg. 1). A GP-based active constraint learning algorithm that iteratively queries start/goal constraint states to induce informative demonstrations that reduce constraint uncertainty.
- High-dimensional evaluation across simulation and hardware. The algorithm is evaluated by inferring high-dimensional nonlinear constraints from generated demonstrations using 4D unicycle and 12D quadcopter dynamics in simulation, and a 7-DOF robot arm hardware platform; the experiments section also lists double integrator dynamics among the simulation settings.
- A decomposition of an intractable active-learning problem. The idealized active constraint learning problem (Prob. 2) is decomposed into sub-problems (Probs. 3-6) that recover start/goal constraint states which are approximately maximally informative, bypassing dependence on the unknown constraint inside the KKT conditions.
The authors also state that, to their knowledge, this provides the first framework for active constraint inference over continuous state spaces.
Main Findings
- Outperforms a random-sampling baseline. Across simulation and hardware experiments using high-dimensional nonlinear dynamics and unknown nonlinear constraints, the method outperforms a random-sampling baseline at accurately performing constraint inference from an iteratively generated set of sparse but informative demonstrations.
- Fewer boundary-approximation errors on the robot arm. In the 7-DOF robot arm result illustrated in Figure 1, GP-ACL recovers a nonlinear obstacle set with fewer false positive (FP) and false negative (FN) errors than the random-sampling baseline.
- Active data generation is the distinguishing capability. Prior constraint-learning methods, including Chou et al. (2022), compute only the mean and covariance functions of a trained GP posterior; this work additionally samples GP posterior samples as surrogate constraint functions to drive the generation of informative demonstrations.
- No requirement for constraint-violating behavior. Unlike the RL-based active constraint learning method of Papadimitriou et al. (2022) and the GP-based method of Li et al. (2025a), the approach can infer unknown high-dimensional constraints in continuous and infinite state and constraint spaces without needing both constraint-satisfying and constraint-violating behavior — which the authors note matters for safety-critical robotics.
- Targets a different inference object than active intent learning. Whereas active intent inference recovers an unknown reward or cost, GP-ACL recovers unknown constraints.
- Quantitative metrics beyond Figure 1's FP/FN comparison are not reported in the available content. The paper content provided is truncated within Section 4.1 (Experiment Setup), before the numerical results tables/figures, so specific error magnitudes, iteration counts, or success rates are not available to report here.
Methodology in Plain English
The method assumes each demonstration is a locally optimal solution to a constrained trajectory optimization problem whose cost penalizes squared consecutive state differences, i.e., trajectory length minimization of the form c(ξ) := Σ_{t=1}^{T-1} ||x_{t+1} − x_t||_2^2. Some constraints are known (dynamics and initial/final state conditions), and one inequality constraint g* is unknown; the unknown constraint is evaluated at a "constraint state" φ_sep(x) obtained from the system state x.
Because each demonstration is locally optimal, it satisfies KKT conditions. At time steps where the unknown constraint is tight (active at equality), the KKT stationarity condition can be used to extract both the constraint state and an estimate of the constraint gradient with respect to the system state. These tight states, their zero constraint values, and the estimated gradients are collected into training sets (D_kappa, D_nabla, D_g, with N_tight = Σ_d |t_tight(ξ_d)|), and a Gaussian process is trained on them, with hyperparameters optimized by marginal log likelihood (Rasmussen and Williams 2006). Plausible constraint functions are drawn from the posterior using a random Fourier feature-based approach.
The idealized active learning objective is to pick start/goal constraint states whose induced demonstrations maximize the GP posterior covariance evaluated at their tight states, subject to the KKT conditions. This is intractable directly, because the KKT conditions involve the unknown constraint and because finding tight time steps requires an inner optimization (Prob. 7). The workaround: draw P posterior samples ĝ_p, and for each one find the constraint state κ_MI,p that maximizes posterior covariance while lying on the sampled constraint boundary (ĝ_p(κ) = 0). Two search schemes are then run in parallel, because the avoid set A = {κ : g*(κ) > 0} may or may not be locally convex near κ_MI,p: one searches within a hyperplane orthogonal to ∇ĝ_p(κ_MI,p) (the locally convex case, Probs. 4-5), and one searches along ±∇ĝ_p(κ_MI,p) (the non-convex case, Prob. 6). Both schemes select start/goal states as close as possible to the offset point κ_p(η) = κ_MI,p + η∇ĝ_p(κ_MI,p) while satisfying two safety conditions: being safe with respect to the sampled constraint by a margin δ, and being safe with respect to the GP posterior with probability at least β, where δ and β are design parameters. These problems are solved with IPOPT (Wächter and Biegler 2006) via Casadi (Andersson et al. 2019), with the safety constraints often encoded as large-weight penalties because initializations may not satisfy them.
Each iteration of GP-ACL trains the GP, draws posterior samples, computes κ_MI,p, generates both orthogonal and parallel start/goal pairs, obtains new demonstrations by solving the demonstrator's forward problem with those start/goal constraints enforced, and appends the new demonstrations and their extracted constraint values to the dataset. Inputs to the algorithm are N_iters, α, n_ℓ, η, an initial demonstration set S_D^(1), δ, and β.
Why This Matters
Impact on research. The paper moves constraint inference from passive to active data collection in the LfD/IOC setting. It relaxes a standing assumption that demonstrations are supplied independently of the downstream learning objective, and it extends active learning — previously explored mainly for cost/reward/intent inference — to constraints over continuous state spaces.
Real-world applications.
- Safety-critical robot manipulation, where a 7-DOF arm must learn an unknown obstacle or keep-out region from a small number of expert demonstrations.
- Aerial robotics, where unknown nonlinear constraints must be inferred for 12D quadcopter dynamics.
- Mobile robot navigation, where constraints on a unicycle-type platform are learned from sparse expert paths.
- Shared human-robot workspaces, where inferring an operator's implicit constraints reduces the need for exhaustive or unsafe trial-and-error.
Industry relevance. Reducing the number of demonstrations needed to characterize an unknown constraint lowers data-collection cost and human time, and generating informative demonstrations through queries rather than exhaustive exploration is directly relevant to deployment scenarios where violating a constraint (for example, colliding) is unacceptable. The stated ability to avoid relying on constraint-violating behavior is a practical advantage for industrial robots operating near people or expensive equipment.
Future Directions
- Extending the formulation beyond state-dependent constraints. The paper notes that the formulation generalizes to settings where some unknown constraints depend on control inputs, which remains to be explored.
- Scaling the active query selection. The idealized problem (Prob. 2) is decomposed into sub-problems and only solved approximately; closing the gap between the approximate solution and the idealized objective is an open question.
- Quantitative characterization of sample efficiency. The abstract and introduction motivate the work by data inefficiency in prior methods, but the available content does not report the specific number of demonstrations or iterations required, which future evaluations would need to quantify.
- Broader comparison against other active constraint learning methods. The paper compares against a random-sampling baseline; comparisons against the RL-based (Papadimitriou et al. 2022) and GP-based (Li et al. 2025a) active constraint learning methods are framed in the related work as differences in capability rather than as empirical benchmarks in the available content.
Target Audience
Robotics researchers working on learning from demonstration, inverse optimal control, and constraint inference; Gaussian process and Bayesian optimization researchers interested in active data acquisition; and safety engineers applying learning-based constraint models to manipulators, quadcopters, or other high-dimensional robotic systems. Readers should be comfortable with KKT conditions, GP regression, and constrained trajectory optimization; the paper is not an introductory treatment.
Authors’ abstract
We present an iterative active constraint learning (ACL) algorithm, within the learning from demonstrations (LfD) paradigm, which intelligently solicits informative demonstration trajectories for inferring an unknown constraint in the demonstrator's environment. Our approach iteratively trains a Gaussian process (GP) on the available demonstration dataset to represent the unknown constraints, uses the resulting GP posterior to query start/goal states, and generates informative demonstrations which are added to the dataset. Across simulation and hardware experiments using high-dimensional nonlinear dynamics and unknown nonlinear constraints, our method outperforms a baseline, random-sampling based method at accurately performing constraint inference from an iteratively generated set of sparse but informative demonstrations.