Technical Summary: AnchorRoute

Problem Statement

Human motion authoring is inherently sparse; users often provide only a few root waypoints, planar trajectory samples, or body-point targets alongside a text prompt. These inputs express intent but do not define a complete motion, leaving the system to infer timing, coordination, and full-body details. Existing sparse-control methods often struggle to balance strict adherence to these sparse anchors with the preservation of high-quality, text-aligned motion generation. The core challenge is to create a framework where sparse anchors can structure both the conditioning of a generative prior and the subsequent correction of the generated motion without degrading the underlying motion quality.

Methodology

AnchorRoute proposes a two-stage framework that utilizes a shared "anchor scaffold" derived from sparse inputs to guide both generation and refinement.

1. The Anchor Scaffold

The framework treats sparse anchors (time, controlled subspace, and target value) as a structural scaffold. This scaffold is instantiated in two ways:

• Generation: As token-aligned condition memory injected into a frozen generator.
• Refinement: As a residual scaffold defining temporal intervals and activation levels for correction.

2. Stage I: Controlled Generation

AnchorRoute builds upon a frozen Transition Masked Diffusion (TMD) text-to-motion prior.

• Dual-Context Conditioning: The system separates text semantics from spatial control. Text provides global action context, while anchors provide sparse spatial control.
• AnchorKV: Sparse anchors are converted into frame-level features (values, masks, interpolation priors, and first-order differences). These are encoded into token-aligned memory ($H_s$) and injected into the frozen TMD transformer via AnchorKV. This mechanism appends anchor-conditioned keys and values to the attention layers, allowing motion tokens to access spatial constraints at every layer without retraining the backbone.
• Training: Only the anchor condition path (scaffold encoder, AnchorKV projections, and dual-context layers) is trained. The TMD backbone remains frozen, preserving the pre-trained text-to-motion generation capabilities.

3. Stage II: RouteSolver Refinement

After the initial motion ($\hat{x}$) is generated, RouteSolver performs inference-time refinement to improve anchor adherence.

• Residual Scaffold: The generated motion is evaluated against the original anchors to compute residuals. These residuals define which temporal intervals require correction and the intensity of that correction.
• Soft-Token Optimization: RouteSolver optimizes a continuous soft-token variable ($u$) in the VQ embedding space. The objective minimizes anchor error, smoothness, and a trust-region penalty to stay close to the initial generation.
• Interval-Routed Projection: Raw optimizer updates are projected onto a piecewise-affine interval basis defined by the anchor timestamps. Crucially, the projection is weighted by residual activities: intervals with larger anchor errors receive more correction capacity, while those with small errors are softly suppressed. This ensures corrections are routed precisely where needed, maintaining coherence.

Key Contributions

1. Anchor Scaffold: A unified control structure that defines both generation-time condition memory and refinement-time interval update spaces using the same sparse anchor inputs.
2. AnchorRoute Framework: A sparse-anchor synthesis system that preserves a frozen TMD prior while learning a lightweight, AnchorKV-based conditioning path with dual-context separation.
3. RouteSolver: A post-generation refinement module that routes soft-token updates through anchor-defined piecewise-affine intervals, guided by residual activities to balance motion quality and anchor adherence.

Experimental Results

Evaluated on the HumanML3D dataset, AnchorRoute demonstrates superior performance in sparse-control tasks:

• Benchmark Comparison: Under the SFControl six-point protocol, AnchorRoute with RouteSolver (RS200) achieves a Control Error of 0.019, significantly outperforming the prior state-of-the-art SFControl (0.036). It also improves FID (0.115 vs. 0.224) and Top-3 R-Precision (0.792 vs. 0.673).
• Control Families: The framework supports Root-3D, Planar-root, and Body-point control within the same formulation. Across all families, increasing refinement steps (RS100 to RS500) progressively reduces Control Error while maintaining a controllable trade-off with motion quality (FID and R-Precision).
• Ablation Studies:
  • Generation: The AnchorKV approach (1.2M trainable params) significantly outperforms Cross-Attention and ControlNet baselines in both Control Error and semantic alignment.
  • Refinement: RouteSolver's use of residual activity routing and piecewise-affine bases yields better quality-control trade-offs compared to plain soft-token refinement or uniform interval updates.
• Efficiency: The default RS200 setting adds approximately 0.093 seconds per sample, offering a practical balance between adherence and inference cost.

Significance and Claims

The paper claims that AnchorRoute successfully decouples the preservation of text-motion quality from the enforcement of sparse spatial constraints. By using a frozen prior for generation and a learned, residual-routed refinement for correction, the system achieves a complementary relationship: the generator maintains high-fidelity motion semantics, while RouteSolver provides a controllable path toward stronger anchor adherence. The work establishes that sparse anchors can effectively structure both the conditioning and the correction phases of motion synthesis, offering a unified solution for diverse control families (root trajectories, planar paths, and body points) without compromising the underlying generative prior.

Limitations: The authors note that RouteSolver currently relies primarily on positional residuals. Future work could extend the residual scaffold to include tangent or orientation cues for direction-aware refinement.