Research map / 3D vision / Human pose estimation
WANDR: Intention-guided Human Motion Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WANDR, a conditional VAE that generates autoregressive human motions reaching arbitrary 3D wrist goals, guided by novel intention features and trained on AMASS and CIRCLE.
Problem
Generating natural human motions in which an avatar walks and reaches the wrist to arbitrary 3D goals, given only an initial pose and goal position; existing RL and data-driven methods generalize poorly and lack naturalness due to scarce goal-reaching training data.
Approach
Intention features (wrist velocity toward goal, body orientation, pelvis-to-goal offset) condition a frame-by-frame c-VAE on pose deltas; hindsight-style pseudo goals derived from future wrist positions let unlabeled AMASS motions be trained alongside goal-labeled CIRCLE data under one objective.
Key results
- WANDR achieves 32% success rate with 16% foot skating and 24.8 cm distance to goal, versus 0% success for GOAL.
- Training on AMASS alone gives 16% SR and 19% FS; CIRCLE alone gives 0% SR with 56% FS, showing the benefit of combining datasets.
- Removing wrist/pelvis/orientation intention components lowers SR to 15-19%; optimization baseline (VAE + opt) reaches 3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on both AMASS and CIRCLE instead of one dataset (ablation) | WANDR · AMASS+CIRCLE eval | Success Rate: 32% vs 16% (AMASS only) and 0% (CIRCLE only) | Helped |
| train only on CIRCLE (no AMASS) (ablation) | WANDR · AMASS+CIRCLE eval | Foot Skating: 56% vs 16% for combined training | Hurt |
| train only on AMASS (no CIRCLE) (ablation) | WANDR · AMASS+CIRCLE eval | Foot Skating: 19% vs 16% for combined (AMASS-only is slightly worse) | Helped |
| add pelvis intention to wrist intention (ablation) | WANDR · AMASS+CIRCLE eval | Success Rate: 18% vs 15% (Iw only) | Helped |
| add orientation intention to wrist intention (ablation) | WANDR · AMASS+CIRCLE eval | Success Rate: 19% vs 15% (Iw only) | Helped |
| add pelvis and orientation intention to wrist intention (full intention) (ablation) | WANDR · AMASS+CIRCLE eval | Success Rate: 32% vs 15% (Iw only); DTG 24.8 vs 62.9 cm | Helped |
| add pelvis and orientation intention to wrist intention (full intention) (ablation) | WANDR · AMASS+CIRCLE eval | Foot Skating: 16% vs 13% (Iw only) | Hurt |
| remove intention and optimize latent space to reach goal (VAE + opt) (ablation) | WANDR · AMASS+CIRCLE eval | Success Rate: 3% vs 32% | Hurt |
| WANDR (full, AMASS+CIRCLE) vs GOAL baseline | WANDR · AMASS+CIRCLE eval | Success Rate: 32% vs 0% | Helped |
Surprising findings
- Training only on CIRCLE yields very high foot skating (56%), indicating the goal-reaching dataset alone cannot teach basic locomotion.
- Wrist-only intention gives the lowest foot skating (13%) but poor success, while adding pelvis and orientation intention increases foot skating.
Limitations
- Error accumulation in autoregressive generation can leave the avatar in unrecoverable states.
- Less proficient at reaching extremely low or high goals due to limited training body-movement diversity.
- No object grasping or scene interaction is modeled.
Experimental setup
Datasets: AMASS CIRCLE
Benchmarks: AMASS held-out split CIRCLE
Metrics: Success Rate Foot Skating Distance to Goal
Baselines: GOAL VAE + opt
Builds on: MotionVAE CIRCLE Hindsight Experience Replay SMPL-X
Closest papers by meaning
- MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction (NeurIPS 2026 · similarity 0.86)
- EnerVerse: Envisioning Embodied Future Space for Robotics Manipulation (NeurIPS 2025 · similarity 0.86)
- NIFTY: Neural Object Interaction Fields for Guided Human Motion Synthesis (CVPR 2024 · similarity 0.85)
- 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · similarity 0.85)
- SemGeoMo: Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance (CVPR 2025 · similarity 0.85)
- Seamless Human Motion Composition with Blended Positional Encodings (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.