Spatial Anchor-based Motion Policy (SAMP)

• Spatial Anchor-based Motion Policy (SAMP) is a design pattern where actions are generated relative to spatially meaningful anchors, ensuring geometry-preserved control.
• It employs diverse anchoring methods—from pixel-aligned maps and fixed initial frames to sparse posture constraints and shared 3D grids—to enhance credit assignment and sample efficiency.
• Empirical evaluations across mobile manipulation, vision-language tasks, and collision-aware planning demonstrate improved success rates and reduced errors compared to traditional methods.

Searching arXiv for the specified papers to ground the article and citations. I’ll look up the listed arXiv entries and closely related SAMP papers to verify bibliographic details. Spatial Anchor-based Motion Policy (SAMP) denotes a class of motion-policy formulations in which action generation is explicitly tied to spatial anchors rather than relying only on an unstructured latent state. Across the literature, the anchor can be a pixel-aligned endpoint in a bird’s-eye map, a persistent initial RGB frame, a sparse set of anchor postures indexed in time, or a shared 3D grid on which robot and environment geometry are co-registered. The resulting policies span mobile manipulation, vision-language-action control, human motion generation, and collision-aware manipulator planning. Taken together, these formulations suggest that SAMP is best understood as a geometry-aligned design pattern whose central operation is to preserve or expose referential structure during action selection or motion synthesis (Wu et al., 2020, Xi et al., 23 Apr 2025, Chen et al., 14 Sep 2025, Zhu et al., 13 Mar 2026).

1. Conceptual scope and terminology

The common thread in SAMP is spatial anchoring: policy outputs are parameterized relative to a reference that is itself spatially meaningful. In the mobile-manipulation formulation grounded in “Spatial Action Maps for Mobile Manipulation,” the anchor is a pixel in a local, robot-aligned bird’s-eye crop, and each pixel denotes a navigational endpoint (Wu et al., 2020). In AnchorVLA4D, the anchor is the first frame of the episode, $I_0$, cached and reused at every timestep to preserve initial scene context (Zhu et al., 13 Mar 2026). In ProMoGen, anchors are sparse posture constraints $A=\{(t_i,x_s^i)\}_{i=1}^{M}$ that specify local joint-space content without displacement, while the trajectory $\tau \in \mathbb{R}^{N\times 3}$ provides global direction and translation (Xi et al., 23 Apr 2025). In the collision-aware manipulator formulation explicitly named SAMP, anchors are uniformly distributed 3D points on a shared workspace grid used to align environment and robot signed distance fields (SDFs) (Chen et al., 14 Sep 2025).

Formulation	Anchor definition	Policy output
Spatial action maps	Pixel $(u,v)$ in a robot-aligned state image	Local navigational endpoint
AnchorVLA4D	Fixed first frame $I_0$	Action chunk for manipulation
ProMoGen	Sparse anchor frames $(t_i,x_s^i)$	Full motion sequence $\hat{X}$
Collision-aware SAMP	Shared 3D spatial grid anchors	Joint increment $\Delta q_{t+1}$

This breadth creates a common misconception: SAMP is not a single canonical algorithm. The literature instead uses the term for several architectures that share a structural commitment to spatially anchored control. A plausible implication is that the term describes a policy family unified by spatial correspondence, not by a single backbone, loss, or action parameterization.

2. Anchor representations and the preservation of context

In the mobile-manipulation setting, the state image is spatially aligned by inverse perspective mapping and local cropping. The policy operates on $I \in \mathbb{R}^{H\times W\times C}$ with $C=4$ channels: an overhead reconstruction map, a robot-position mask, shortest-path distance-from-agent, and shortest-path distance-to-receptacle. The action representation is a dense Q-value map $A=\{(t_i,x_s^i)\}_{i=1}^{M}$0, where each pixel $A=\{(t_i,x_s^i)\}_{i=1}^{M}$1 corresponds to the action “navigate to local endpoint at $A=\{(t_i,x_s^i)\}_{i=1}^{M}$2.” Because $A=\{(t_i,x_s^i)\}_{i=1}^{M}$3 is aligned with $A=\{(t_i,x_s^i)\}_{i=1}^{M}$4, actions are spatially anchored to local visual structure, which the formulation attributes to improved credit assignment, translational equivariance, and sample efficiency (Wu et al., 2020).

AnchorVLA4D reinterprets anchoring as episodic memory for robot manipulation. The anchor image is defined as the first frame of the episode, $A=\{(t_i,x_s^i)\}_{i=1}^{M}$5, and it remains fixed throughout execution. Its role is explicitly twofold: temporal memory and spatial stabilization. By conditioning each action on both $A=\{(t_i,x_s^i)\}_{i=1}^{M}$6 and $A=\{(t_i,x_s^i)\}_{i=1}^{M}$7, the policy preserves target geometry when the current view becomes partially occluded and reduces spatial disorientation during retries. The formulation states that anchoring causes earlier retries after failure and prevents repeating the same mistaken grasp location, while requiring no explicit recurrent state and no depth or point clouds (Zhu et al., 13 Mar 2026).

In ProMoGen, anchoring is temporal but not episodic in the same sense. Sparse anchor postures $A=\{(t_i,x_s^i)\}_{i=1}^{M}$8 are displacement-free action constraints; they specify local posture details at selected frames, while the global trajectory $A=\{(t_i,x_s^i)\}_{i=1}^{M}$9 governs pelvis translation. This separation is both representational and architectural. The anchor set constrains motion locally in joint space, and the trajectory provides the global path. The formulation therefore treats spatial anchoring as selective control over motion content rather than as a memory cue (Xi et al., 23 Apr 2025).

In the collision-aware manipulator formulation, anchors are neither images nor discrete postures but shared spatial samples. Environment SDF features and robot SDF features are both evaluated on a common grid, which permits direct geometric correspondence at the points where collision reasoning matters. This shifts the notion of anchoring from temporal persistence to feature alignment in 3D workspace coordinates (Chen et al., 14 Sep 2025).

3. Policy formulations and learning objectives

The mobile-manipulation formulation defines the policy directly on the spatial map:

$\tau \in \mathbb{R}^{N\times 3}$0

A stochastic variant is also given via a spatial softmax over pixels. Learning uses Double DQN with smooth L1 (Huber) loss and a target network, and the effective step discount is scaled by traveled distance,

$\tau \in \mathbb{R}^{N\times 3}$1

This distance-dependent discount is introduced to avoid over-crediting long moves under variable-length actions (Wu et al., 2020).

AnchorVLA4D defines an anchor-conditioned manipulation policy

$\tau \in \mathbb{R}^{N\times 3}$2

where $\tau \in \mathbb{R}^{N\times 3}$3 fuses visual-language features with spatial-encoder features. The action is continuous, with

$\tau \in \mathbb{R}^{N\times 3}$4

where $\tau \in \mathbb{R}^{N\times 3}$5, $\tau \in \mathbb{R}^{N\times 3}$6, and the gripper is binary or continuous. The action expert is a DDPM-style conditional diffusion model over action chunks, trained with noise-prediction MSE rather than a direct supervised MSE on actions:

$\tau \in \mathbb{R}^{N\times 3}$7

The model jointly predicts the next $\tau \in \mathbb{R}^{N\times 3}$8 actions during training (Zhu et al., 13 Mar 2026).

ProMoGen casts SAMP as conditional motion generation:

$\tau \in \mathbb{R}^{N\times 3}$9

Its diffusion process combines a trajectory encoder, an anchor-motion encoder, an initial generator, a refinement module, and a decoder. The total objective is

$(u,v)$0

with $(u,v)$1 and $(u,v)$2. The terms enforce reconstruction, anchor adherence, local joint accuracy, adversarial naturalness, and physics constraints (Xi et al., 23 Apr 2025).

The collision-aware manipulator formulation predicts joint increments rather than full trajectories in one shot. It defines environment and robot geometry by signed distance fields, with

$(u,v)$3

and trains a policy using point matching, collision, and configuration losses:

$(u,v)$4

In the reported experiments, $(u,v)$5, $(u,v)$6, and $(u,v)$7 are used as the core terms (Chen et al., 14 Sep 2025).

4. Architectural realizations

The architectural realization of SAMP varies with the anchor domain, but each version enforces alignment between anchors and action generation. In the spatial-action-map setting, a modified fully convolutional ResNet-18 removes AvgPool, the fully connected classifier, and all BatchNorm layers, and appends a 3-layer $(u,v)$8 convolutional head with bilinear upsampling to recover a dense $(u,v)$9 Q-map. The selected pixel is mapped to a local Cartesian waypoint $I_0$0 via the calibrated inverse perspective mapping, and the robot executes either a straight-line primitive or a shortest-path primitive (Wu et al., 2020).

AnchorVLA4D combines Qwen2.5-VL (3B), Any4D as a lightweight spatial encoder, and ScaleDP (400M) as the diffusion-based action head. The visual-language backbone encodes $I_0$1 and $I_0$2 and fuses them with the instruction $I_0$3 to produce $I_0$4, while the spatial encoder jointly processes $I_0$5 and $I_0$6 to produce $I_0$7. The best-performing fusion is simple concatenation before the action expert, followed by adaptive pooling or projection to the conditioning dimension. Alternative injections were inferior: cross-attention before the VL decoder yielded $I_0$8 success rate, injection in the action head yielded $I_0$9, and concatenation achieved $(t_i,x_s^i)$0. The spatial encoder is pretrained and, in the best configuration, kept frozen; unfreezing it without 3D supervision degraded performance by $(t_i,x_s^i)$1 (Zhu et al., 13 Mar 2026).

ProMoGen uses a hierarchical transformer-based diffusion architecture. The trajectory encoder $(t_i,x_s^i)$2 and anchor-motion encoder $(t_i,x_s^i)$3 are decoupled, and their outputs are fused through an initial motion generator and a refinement module:

$(t_i,x_s^i)$4

Sampling is accelerated by DPM-Solver++ with 25 inference steps. A distinctive component is the Spatial Anchor Filtering Module, which simulates user-specified sparse anchors during training under explicit controls on temporal density and minimum spacing, and SAP-CL then varies those controls in an easy-to-hard curriculum (Xi et al., 23 Apr 2025).

The collision-aware manipulator architecture uses a 3D CNN over the environment SDF grid, a dedicated robot SDF network based on a multiresolution hash-grid encoder and MLP decoder, a PointNet-style encoder for robot sample points, and MLP heads for configuration encoding, feature fusion, and action decoding. Feature alignment is performed by joint grid sampling: environment features are trilinearly interpolated at transformed robot sample points, then fused with robot SDF features sampled at the same locations. This gives a stepwise policy in which geometry is co-registered before joint increments are predicted (Chen et al., 14 Sep 2025).

A recurrent theme across these systems is that simple alignment mechanisms can outperform heavier temporal or cross-attentional alternatives when the anchor is already well matched to the task geometry. This suggests that, within SAMP, the quality of anchoring can matter more than architectural complexity.

5. Empirical performance and observed behavior

The strongest quantitative evidence for SAMP in vision-language manipulation comes from AnchorVLA4D. On the Simpler WidowX benchmark, AnchorVLA4D reaches a $(t_i,x_s^i)$5 success rate, a $(t_i,x_s^i)$6 percentage point improvement over VanillaVLA at $(t_i,x_s^i)$7. AnchorVLA alone improves success rate by $(t_i,x_s^i)$8 over VanillaVLA, and the spatial encoder adds a further $(t_i,x_s^i)$9 on top of AnchorVLA, from $\hat{X}$0 to $\hat{X}$1. In real-world tasks on XLeRobot, the reported average success rate across “Lift box with two arms,” “Open drawer,” and “Rotation and pour” is $\hat{X}$2, exceeding the $\hat{X}$3 baseline at $\hat{X}$4. The same study reports that replacing the single anchor with the previous 3 frames yields $\hat{X}$5 success rate, and using frames at stride 20 yields $\hat{X}$6, indicating that a single persistent anchor can outperform heavier temporal stacks under the evaluated conditions (Zhu et al., 13 Mar 2026).

The original spatial-action-map formulation reports large improvements over low-dimensional steering baselines in navigate-and-push tasks. With the shortest-path primitive, mean objects per episode over five runs were $\hat{X}$7 on SmallEmpty, $\hat{X}$8 on SmallColumns, $\hat{X}$9 on LargeColumns, and $\Delta q_{t+1}$0 on LargeDivider. The 18-action steering baseline achieved $\Delta q_{t+1}$1, $\Delta q_{t+1}$2, $\Delta q_{t+1}$3, and $\Delta q_{t+1}$4 on the same environments. In the real world, the robot pushes $\Delta q_{t+1}$5 objects within 15 minutes and all 10 within 30 minutes on SmallEmpty (Wu et al., 2020).

The collision-aware manipulator formulation reports an overall $\Delta q_{t+1}$6 increase in success rate and $\Delta q_{t+1}$7 reduction in collision rate relative to learning-based baselines. A representative configuration with 160 anchors achieves success of $\Delta q_{t+1}$8 on train and $\Delta q_{t+1}$9 on test, collision rates of $I \in \mathbb{R}^{H\times W\times C}$0 on train and $I \in \mathbb{R}^{H\times W\times C}$1 on test, and solution time of approximately $I \in \mathbb{R}^{H\times W\times C}$2–$I \in \mathbb{R}^{H\times W\times C}$3 s per trajectory. The dedicated robot SDF model attains MAE $I \in \mathbb{R}^{H\times W\times C}$4 mm and RMSE $I \in \mathbb{R}^{H\times W\times C}$5 mm, with training time of about $I \in \mathbb{R}^{H\times W\times C}$6 s and inference time of about $I \in \mathbb{R}^{H\times W\times C}$7 ms (Chen et al., 14 Sep 2025).

ProMoGen’s evaluation is reported on HumanML3D and CombatMotion using MPJPE, joint smoothness, Diversity, Directional Consistency, K-MPJPE, and FID. Ours-v2 is reported as achieving the best overall performance across metrics against baselines including MDM, MotionDiffuse, StableMoFusion, and OmniControl. SAP-CL improves performance relative to regular training across anchor counts $I \in \mathbb{R}^{H\times W\times C}$8-Num $I \in \mathbb{R}^{H\times W\times C}$9, and ablations identify the Trajectory Feature Encoder as pivotal, followed by the Initial Motion Generator and Anchor Motion Encoder (Xi et al., 23 Apr 2025).

Formulation	Setting	Reported result
AnchorVLA4D	Simpler WidowX	$C=4$0 success rate
Spatial action maps	LargeDivider	$C=4$1 objects/episode
Collision-aware SAMP	Test environments, 160 anchors	$C=4$2 success, $C=4$3 collision
ProMoGen	HumanML3D and CombatMotion	Best overall across reported metrics

6. Limitations, misconceptions, and research directions

Each SAMP formulation inherits limitations from its anchor design. In AnchorVLA4D, the anchor is fixed and never updated. This creates anchor staleness when the execution state drifts far from the initial scene, particularly in long-horizon tasks or when large gripper rotations are required. The study explicitly suggests incremental anchor updates as a potential remedy but does not implement them. Severe occlusion, dynamic scene changes, and ambiguous language remain failure modes (Zhu et al., 13 Mar 2026).

The spatial-action-map formulation depends on accurate spatial alignment and mapping. Misalignment or SLAM errors can mis-anchor actions, unseen obstacles treated as free space can produce collisions, and the single-channel endpoint map is best suited to “navigate-to-and-push” rather than rich manipulation involving orientation, grasp pose, or continuous force control (Wu et al., 2020).

ProMoGen remains sensitive to extreme anchor sparsity. Very sparse anchors, such as $C=4$4-Num $C=4$5, degrade precision relative to denser settings even with SAP-CL, and generalization remains bounded by dataset coverage. The method is oriented toward combining trajectory and sparse anchor constraints rather than high-level semantics (Xi et al., 23 Apr 2025).

The collision-aware manipulator formulation trades geometric fidelity for compute. A high-resolution 3D CNN over dense SDF grids increases runtime, extremely tight spaces can leave no feasible corridor under the chosen safety margin, and dynamic obstacles are not explicitly handled. The method also depends on SDF accuracy; poor map quality can mislead the policy (Chen et al., 14 Sep 2025).

A further misconception is that spatial anchoring necessarily requires additional sensing modalities or large temporal memories. AnchorVLA4D explicitly reports gains without depth or point clouds and with negligible inference overhead in simulation: baseline latency is $C=4$6 s, while anchor plus spatial encoder raises it to $C=4$7 s, about a $C=4$8 increase, corresponding to approximately $C=4$9 Hz effective control (Zhu et al., 13 Mar 2026). Conversely, the manipulator SAMP formulation shows that anchoring can be entirely geometric, with anchors instantiated as shared 3D grid points rather than images or temporal cues (Chen et al., 14 Sep 2025).

Taken together, these results suggest that the enduring significance of SAMP lies less in any particular implementation than in a recurring methodological principle: motion policies become more stable and more controllable when their decision variables are anchored to explicit spatial references that preserve geometry, temporal context, or both.