Research map / Probabilistic and scientific modeling / Molecules and proteins

Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment

CVPR 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Multimodal trajectory prediction · 8 citations (OpenAlex)

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 Locomotion Embodiment, which trains a differentiable Locomotion Value (LocoVal) surrogate of a physics-based humanoid locomotion generator (PACER) and uses it as an EmLoco loss to regularize trajectory predictors and as a filter at inference.

Problem

Human trajectory prediction (HTP) from momentary observations often produces physically implausible trajectories, because pose cues are used only implicitly and indifferentiable physics simulators cannot supervise HTP networks directly.

Approach

Distill the plausibility of physics-simulated humanoid locomotion (discounted reward of a trajectory-following policy) into a differentiable MLP (LocoVal), then use it as a regularization loss on all predicted heads of a multi-head HTP network and as an inference-time filter.

Key results

• On JTA with standard 9-frame setting, the method reaches ADE 0.97 / FDE 1.91 vs Social-Trans 1.11 / 2.26.
• With 2-frame momentary observations on JTA (deterministic), ADE 1.33 vs Social-Trans 1.36; with 5 heads minADE 0.87 vs 0.92.
• LocoVal filter on zero-shot ETH/UCY with EqMotion reduces mean ADE from 1.12 to 0.88 and FDE from 2.41 to 1.89.

Measured effects reported in the paper

Surprising findings

• Training with only the EmLoco loss (no ground-truth MSE/minMSE loss) still improves over an randomly initialized network, with JTA ADE going from 5.07 to 3.52.
• Filtering rejects trajectories that are much worse, and the LocoVal function estimates plausibility from root velocity alone well enough to help even without poses on ETH/UCY.

Limitations

• Performance gains are smaller on JRDB, attributed to noisy estimated 3D poses that the method assumes to be accurate.
• Requires accurate 3D poses from AMASS and a physics-based locomotion generator to train the LocoVal function.
• Some JTA trajectories are infeasible for a humanoid and are deprioritized, which may limit coverage.

Experimental setup

Datasets: JTA JRDB AMASS ETH/UCY

Benchmarks: JTA JRDB ETH/UCY

Metrics: ADE FDE minADE minFDE chi-squared distance over velocity, acceleration, angular ve

Baselines: Social-Transmotion Social-GAN Social-LSTM Trajectron++ Autobots EqMotion HST MATRIX

Builds on: Social-Transmotion PACER minMSE multi-head prediction MATRIX

Closest papers by meaning

• Human Motion Prediction Under Unexpected Perturbation (CVPR 2024 · similarity 0.86)
• Universal Humanoid Motion Representations for Physics-Based Control (ICLR 2024 · similarity 0.86)
• SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.86)
• MoFlow: One-Step Flow Matching for Human Trajectory Forecasting via Implicit Maximum Likelihood Estimation based Distillation (CVPR 2025 · similarity 0.85)
• Multi-modal Knowledge Distillation-based Human Trajectory Forecasting (CVPR 2025 · similarity 0.85)
• HVIS: A Human-like Vision and Inference System for Human Motion Prediction (AAAI 2025 · similarity 0.85)