Sparse Goal Candidate Proposal (SGCP)

• SGCP is a methodological pattern that generates a sparse, task-relevant set of candidate goals or hypotheses prior to final selection in planning and control.
• It integrates four roles—candidate generation, sparsification, feasibility filtering, and downstream consumption—across diverse areas like trajectory prediction, reinforcement learning, and inverse kinematics.
• SGCP methods improve performance by reducing search space while preserving operational properties such as transferability, feasibility, and planner alignment.

Sparse Goal Candidate Proposal (SGCP) denotes a class of methods that construct a small, task-relevant set of candidate goals, subgoals, endpoints, or target hypotheses before downstream selection, planning, or control. The term appears explicitly in "BEVTraj: Map-Free End-to-End Trajectory Prediction in Bird's-Eye View with Deformable Attention and Sparse Goal Proposals" (Kong et al., 12 Sep 2025), where a decoder predicts a sparse set of $K$ goal candidates from the target agent’s dynamic state and a bird’s-eye-view feature map. Across adjacent literatures, however, closely related mechanisms also appear under other names: transferable candidate proposal for active learning (Go et al., 2023), density-based curriculum over achieved goals (Yang et al., 2021), sparse goal-aware graph representations for generalized planning (Jeon et al., 14 Aug 2025), autonomous goal selection inside hierarchical inverse kinematic optimization (Pfeiffer, 2024), sparse object-centered hypothesis generation for instance image-goal navigation (Narasimhan et al., 17 Nov 2025), and planner-admissible value extension from sparse goal-conditioned labels (Zhang, 18 May 2026). This suggests that SGCP is best treated as a methodological pattern rather than a single canonical algorithm.

1. Conceptual scope and defining structure

A recurring distinction in this literature is the separation between candidate proposal and final selection. In transferable active learning, the proxy model is not asked to identify the exact final subset for the target model; instead, it narrows the unlabeled pool to a transferable candidate set, after which the target model applies its own acquisition rule on that reduced pool (Go et al., 2023). In multi-goal reinforcement learning, achieved goals are first sampled from low-density regions and then filtered or ranked by desired-goal density before being used for exploration and replay (Yang et al., 2021). In map-free trajectory prediction, SGCP explicitly predicts a compact set of goal endpoints that are consumed by downstream trajectory decoders rather than by a separate post-processing stage (Kong et al., 12 Sep 2025).

A second recurring property is that sparsity is usually not imposed as an abstract compression objective alone. Instead, the sparse set is intended to preserve some operational property: transferability across model mismatch, feasibility under sparse rewards, kinematic realizability, planner admissibility, or scene consistency. In the inverse-kinematic setting, sparse selection is embedded directly inside a hierarchical nonlinear program, so that goal choice is solved simultaneously with full inverse kinematics and task priorities rather than by a separate reachability approximation (Pfeiffer, 2024). In graph-based sparse goal-conditioned planning, sparse labels on a goal-dependent boundary are extended so that a greedy planner reaches the goal, making the relevant notion of quality operational rather than purely interpolative (Zhang, 18 May 2026).

This suggests a common SGCP decomposition with four roles: a candidate generator, a sparsifier or ranker, a feasibility or transfer filter, and a downstream consumer. The downstream consumer differs sharply by domain—active-learning acquisition, episode-goal replacement, trajectory decoding, inverse-kinematic optimization, frontier exploration, or greedy graph planning—but the upstream structural idea is stable: reduce a large search space to a smaller set of plausible goal-relevant hypotheses.

2. Representative instantiations across domains

The literature instantiates SGCP-like mechanisms in several technically distinct ways.

Domain	Candidate source	Downstream consumer
Active learning	$C = U_x \setminus (S_{LE}\cup S_{HA})$	Acquisition function $\mathcal{A}$ on $C$
Multi-goal RL	Low-density achieved goals and local augmentations	Episode-goal replacement and replay relabeling
Trajectory prediction	$K$ learned goal coordinates in $\mathbb{R}^{K\times 2}$	Initial prediction and iterative refinement
IK planning/control	Candidate Cartesian goals encoded as sparse constraints	Hierarchical nonlinear solver
Image-goal navigation	Object centroids and frontier locations	Matching, frontier scoring, and navigation
Graph planning	Sparse labeled boundary $\Gamma_g$	Greedy argmin-$Q$ rollout

In active learning, the core object is the candidate pool

$C = U_x \setminus (S_{LE} \cup S_{HA}),$

where low-epistemic (LE) and high-aleatoric (HA) unlabeled points are removed before the target-side acquisition step (Go et al., 2023). In multi-goal RL, candidate goals are drawn from the achieved-goal replay buffer with preference for low-density regions,

$$p(ag_i)= \frac{e^{1-\hat{\rho}_i}}{\sum_{n=1}^{N} e^{1-\hat{\rho}_n}},$$

and then re-ranked using desired-goal density entropy,

$C = U_x \setminus (S_{LE}\cup S_{HA})$0

so that exploration expands toward task-relevant frontier regions (Yang et al., 2021).

In BEV trajectory prediction, SGCP is explicit: learnable seed parameters

$C = U_x \setminus (S_{LE}\cup S_{HA})$1

are fused with target dynamics and BEV features, transformed into mode-specific queries $C = U_x \setminus (S_{LE}\cup S_{HA})$2, then into content queries $C = U_x \setminus (S_{LE}\cup S_{HA})$3, and finally regressed to $C = U_x \setminus (S_{LE}\cup S_{HA})$4 goal coordinates in $C = U_x \setminus (S_{LE}\cup S_{HA})$5 (Kong et al., 12 Sep 2025). In inverse kinematics, candidate goals are represented as a selection-constraint group

$C = U_x \setminus (S_{LE}\cup S_{HA})$6

and sparse slack minimization typically makes exactly one candidate feasible for a given end-effector (Pfeiffer, 2024). In image-goal navigation, the candidate set is object-centered,

$C = U_x \setminus (S_{LE}\cup S_{HA})$7

with persistent 3D centroids, supporting Gaussians, and observed viewpoints (Narasimhan et al., 17 Nov 2025). In graph planning, sparse goal-conditioned supervision is carried by a labeled boundary $C = U_x \setminus (S_{LE}\cup S_{HA})$8, and the extension problem is to construct $C = U_x \setminus (S_{LE}\cup S_{HA})$9 on all vertices so that greedy rollouts reach the goal (Zhang, 18 May 2026).

3. Core mechanisms of sparsification, scoring, and conditioning

A large fraction of SGCP-style methods rely on structured pruning rather than unconstrained search. In active learning, Transferable candidate proposal with Bounded Uncertainty uses bounded uncertainty literally: the candidate pool is the middle region between low epistemic uncertainty and persistently low-confidence, high-aleatoric samples (Go et al., 2023). Epistemic uncertainty is estimated with a last-layer Laplace approximation and predictive entropy, while aleatoric screening uses FreeMatch-style adaptive class-wise confidence thresholds over five periodic evaluations. The resulting preselection architecture is strict: TBU does not replace the acquisition function; it shrinks the search space the target-side selector sees.

In multi-goal RL, sparsification operates through density estimation and curriculum. Achieved-goal density $\mathcal{A}$0 identifies under-explored but already achieved frontier points, while desired-goal density $\mathcal{A}$1 filters those points toward task-relevant regions (Yang et al., 2021). This is followed by goal augmentation inside the success-preserving neighborhood

$\mathcal{A}$2

which thickens useful sparse proposals without changing the sparse-reward success label. The proposed goals then enter both online exploration and replay-time relabeling.

In BEVTraj, sparsification is query-based rather than density-based. SGCP predicts a fixed sparse set of endpoint hypotheses directly, supervises them against the ground-truth final destination with

$\mathcal{A}$3

and trains a per-candidate displacement estimate via

$\mathcal{A}$4

Because the model emits only $\mathcal{A}$5 continuous coordinates rather than a dense goal field, the paper states that SGCP enables end-to-end prediction without post-processing such as non-maximum suppression (Kong et al., 12 Sep 2025).

In SplatSearch, the sparse candidate itself is enriched before scoring. Sparse-view 3D Gaussian Splatting renders multiple candidate-centered views, and a multi-view diffusion model completes missing regions in those views. Candidate matching is then performed by selecting the best rendered viewpoint for each object,

$\mathcal{A}$6

with threshold-based acceptance when $\mathcal{A}$7 (Narasimhan et al., 17 Nov 2025). This makes completion an intrinsic part of the candidate-proposal stack rather than a separate recognition refinement stage.

4. Feasibility, transferability, and planner alignment

A central issue in SGCP research is that sparse proposal is only useful when it preserves the right downstream decisions. The active-learning literature makes this explicit by rejecting the assumption that there exists a universally informative subset independent of model configuration. TBU therefore solves a weaker problem than direct subset transfer: it discards points believed to be broadly non-useful across models and leaves the target model the right to choose the actually informative subset under its own acquisition function and architecture (Go et al., 2023). This suggests a general SGCP principle: proposal can be more transferable than final selection when model mismatch is substantial.

The multi-goal RL literature emphasizes feasibility. Density-based curriculum methods select achieved rather than arbitrary imagined goals, so candidate goals are reachable at least once by construction; desired-goal density then constrains novelty toward relevant and safe regions (Yang et al., 2021). The minimum-time sparse-reward literature sharpens this issue further by arguing that sparse goal-reaching learnability depends strongly on early accidental success. In that setting, the goal-hit rate of the initial policy is identified as a robust early indicator for learning success, and the paper reports that achieving an average of $\mathcal{A}$8 target hits per $\mathcal{A}$9 steps is sufficient for successful learning in their SAC setup (Vasan et al., 2024). A plausible implication is that SGCP systems for sparse-reward control should score candidates not only by semantic relevance but also by expected hit probability under the current policy and timeout structure.

Planner alignment appears most explicitly in sparse goal-conditioned graph planning. The operational planner is

$C$0

and the main local certificate states that if the surrogate error along the rollout satisfies

$C$1

then the greedy rollout reaches the goal (Zhang, 18 May 2026). For AMLE, this is instantiated through a fill-distance bound; under the stated assumptions, the condition

$C$2

is sufficient along the rollout. By contrast, harmonic extension can mis-rank local actions because its values reflect boundary hitting probabilities rather than shortest-path greedy order. This sharply distinguishes planner-admissible sparse propagation from merely smooth interpolation.

Inverse-kinematic SGCP provides a parallel notion of feasibility, but inside optimization rather than rollout. Sparse hierarchical nonlinear programming places autonomous goal selection on any priority level, with higher-priority feasibility constraints inherited via active and inactive sets, so lower-priority goal selection occurs only in the remaining feasible motion subspace (Pfeiffer, 2024). In that sense, feasibility is not a post hoc filter; it is built into the proposal mechanism itself.

5. Empirical evidence and domain-specific performance

Empirical results across domains show that sparse candidate proposal can improve the downstream task when the proposal mechanism is aligned with the task’s true decision structure.

In transferable active learning, TBU was verified on CIFAR-10, CIFAR-100, and SVHN. On CIFAR-10, the paper reports that for Entropy in round 2, SAME achieved $C$3, DIFF $C$4, SEMI $C$5, and TBU($C$6) $C$7; for Badge in round 4, SAME was $C$8, DIFF $C$9, and TBU($K$0) $K$1 (Go et al., 2023). The appendix also states that TBU continues to outperform most baselines when the target architecture changes to VGG-16, supporting architecture-transfer relevance.

In density-based curriculum learning for multi-goal RL, the proposed method is evaluated on five sparse-reward Fetch tasks and is reported to outperform HER, CHER, MEP, HGG, and OMEGA. On the hardest task, FetchPnP-Obstacle, HER, CHER, and MEP fail, while the proposed method succeeds; compared to HGG and OMEGA, the paper claims over $K$2 speedup in exploration efficiency in the hardest setting (Yang et al., 2021). The ablations on goal exploration, goal augmentation, and transition augmentation show that removing any component hurts performance.

In sparse minimum-time goal-reaching RL, the paper contrasts dense shaping with a constant negative reward until success and shows that the sparse formulation can yield better final policies. On Reacher-Easy the minimum-time policy is approximately $K$3 faster, and on Reacher-Hard approximately $K$4 faster in steps to goal; it also stays inside the target area longer, resulting in over 500 higher accrued rewards (Vasan et al., 2024). The paper further reports learning pixel-based policies from scratch on four real-robotic platforms within roughly $K$5–$K$6 hours using constant negative rewards.

In hierarchical inverse-kinematic goal selection, the HRP-2Kai planning experiment gives four end-effectors two candidate goals each and reports nearly zero slack for exactly one goal per end-effector, consistent with sparse autonomous selection (Pfeiffer, 2024). The strongest scaling result is the 100-candidate right-hand target experiment: with autonomous goal selection enabled, the robot touches 93 objects before they pass; without autonomous goal selection, it touches none. The proposed $K$7QP solver solves the SHQP in $K$8 ms in most iterations.

In BEVTraj, the SGCP ablation directly isolates sparse goal proposal. Without SGCP, the reported metrics are minADE$K$9 $\mathbb{R}^{K\times 2}$0, minADE$\mathbb{R}^{K\times 2}$1 $\mathbb{R}^{K\times 2}$2, minFDE$\mathbb{R}^{K\times 2}$3 $\mathbb{R}^{K\times 2}$4, minFDE$\mathbb{R}^{K\times 2}$5 $\mathbb{R}^{K\times 2}$6, and Miss Rate $\mathbb{R}^{K\times 2}$7; with SGCP, these improve to $\mathbb{R}^{K\times 2}$8, $\mathbb{R}^{K\times 2}$9, $\Gamma_g$0, $\Gamma_g$1, and $\Gamma_g$2, respectively (Kong et al., 12 Sep 2025). The largest gain is in endpoint error, which is consistent with better goal proposal quality.

In instance image-goal navigation, SplatSearch reports HM3D-val SR $\Gamma_g$3 and SPL $\Gamma_g$4, versus $\Gamma_g$5/$\Gamma_g$6 for UniGoal and $\Gamma_g$7/$\Gamma_g$8 for IEVE; on HM3D-val-hard, SplatSearch reports SR $\Gamma_g$9 and SPL $Q$0 (Narasimhan et al., 17 Nov 2025). Ablations show degradation without novel viewpoint synthesis, without the view-consistent image completion network, and without semantic or visual context scores, supporting the interpretation that sparse candidate generation, candidate completion, and frontier scoring are jointly responsible for performance.

In planner-admissible graph value extension, the aggregate rollout success over 120 AntMaze graph configurations is $Q$1 for harmonic extension and $Q$2 for AMLE; finite high-$Q$3 methods also enter a high-success regime, with success $Q$4 for $Q$5, $Q$6 for $Q$7, and $Q$8 for a fixed-budget $Q$9 solver (Zhang, 18 May 2026). Mechanism audits report that many rollout decisions occur in AMLE-compatible but harmonic-incompatible local geometry, and that AMLE corrects most harmonic inversions on the rollout-weighted decision scope.

6. Limitations, misconceptions, and open problems

A common misconception is that SGCP necessarily refers to an explicit learned goal-proposal block. The literature does not support that reading. Explicit SGCP terminology appears in BEVTraj (Kong et al., 12 Sep 2025), but several technically relevant works are only approximate or implicit instances of the pattern: TBU is a transferable candidate proposal layer for active learning rather than a goal proposer (Go et al., 2023); density-based curriculum is an implicit curriculum-based goal sampler and relabeler rather than a standalone SGCP module (Yang et al., 2021); the sparse goal-aware GNN paper offers sparse, goal-aware representation design but does not implement a true sparse goal candidate proposal module (Jeon et al., 14 Aug 2025); and planner-admissible graph-PDE value extension addresses sparse candidate-goal propagation rather than candidate generation itself (Zhang, 18 May 2026).

A second misconception is that sparsity alone is the relevant design axis. The papers instead couple sparsity with transfer, feasibility, geometry, or admissibility. TBU can over-prune useful points if uncertainty estimation is poorly calibrated, and candidate size is only indirectly controlled by $C = U_x \setminus (S_{LE} \cup S_{HA}),$0 and the HA criterion (Go et al., 2023). Density-based curriculum can only propose near the achieved distribution and depends on KDE in low-dimensional goal spaces (Yang et al., 2021). Minimum-time sparse-reward RL does not provide a goal proposal algorithm, reachable-set estimator, or HER-style replay amplification (Vasan et al., 2024). The sparse goal-aware GNN retains all nodes and uses manual locality-based relation pruning rather than learned candidate selection (Jeon et al., 14 Aug 2025). Inverse-kinematic sparse selection remains local in the usual SQP sense and depends on distinct candidate goals and differentiable task maps (Pfeiffer, 2024). SplatSearch depends on detection and segmentation quality, on the accuracy of the online sparse 3DGS map, and on diffusion-based completion priors (Narasimhan et al., 17 Nov 2025).

The strongest open directions follow directly from these limitations. The graph-PDE literature explicitly identifies adaptive sparse-label selection as an open problem (Zhang, 18 May 2026). The SGCP interpretation of that result is immediate: once sparse candidate proposals are viewed as boundary anchors, candidate placement itself becomes a coverage problem over rollout-relevant neighborhoods. A plausible implication is that future SGCP systems will need to co-design candidate generation, candidate propagation, and downstream planner or controller structure, rather than treating sparse proposal as an isolated front end.