BiMotion: Multi-Modal Motion Planning & Synthesis

• BiMotion is a multifaceted paradigm involving bi-directional motion planning, continuous B-spline motion synthesis for text-guided 3D generation, and bilingual speech emotion recognition benchmarks.
• It employs techniques like lazy dynamic programming, diffusion-transformer models, and cross-lingual embedding alignment to improve efficiency and semantic consistency.
• Empirical results reveal significant runtime reductions in planning, superior animation quality scores, and enhanced performance in cross-group speech emotion recognition tasks.

BiMotion denotes multiple distinct research concepts in recent academic literature, each employing the term toward a specialized methodology for either motion planning, 3D motion generation, or speech emotion recognition. These include: (1) the “BiMotion” principle for bi-directional sampling-based motion planning; (2) BiMotion, a continuous B-spline representation and diffusion framework for text-based 3D character motion; (3) BiMotion benchmarks for bilingual, bi-age-group speech emotion recognition. This entry details all three usages as established in primary arXiv sources.

1. BiMotion in Sampling-Based Bi-directional Motion Planning

The “BiMotion” principle formalizes a bi-directional search strategy for optimal motion planning in continuous spaces, instantiated by the Bi-directional Fast Marching Tree algorithm (BFMT*) (Starek et al., 2015). In this context, the configuration space is a compact subset $X \subset \mathbb{R}^d$ with obstacles $X_{\mathrm{obs}}$ and free space $$X_{\mathrm{free}} = \mathrm{cl}(X \setminus X_{\mathrm{obs}})$$. The planning query entails finding a continuous, collision-free path $\sigma : [0,1] \to X_{\mathrm{free}}$ connecting start $x_{\mathrm{init}}$ to goal $x_{\mathrm{goal}}$, minimizing the cost functional $c(\sigma) = \int_0^1 \|\dot{\sigma}(t)\| dt$.

Distinctively, BiMotion considers two value functions: (a) cost-to-come $V^+(x)$ from $x_{\mathrm{init}}$ and (b) cost-to-go $V^-(x)$ to $X_{\mathrm{obs}}$0, such that the optimal path minimizes $X_{\mathrm{obs}}$1 over some meeting state $X_{\mathrm{obs}}$2. BFMT* implements two simultaneous, lazy dynamic programming expansions (forward and backward trees), interleaving their growth and performing minimal collision checks. A path is returned as soon as a sample is added to both trees, ensuring asymptotic optimality under probabilistic exhaustivity and minimal connection radius scaling ($X_{\mathrm{obs}}$3).

Empirical studies demonstrate that BFMT* achieves similar or better solution cost in 30–50% less runtime than unidirectional FMT, and is $X_{\mathrm{obs}}$4–$X_{\mathrm{obs}}$5 faster than RRT/PRM* in moderate to high dimensions. Particularly, in uncluttered $X_{\mathrm{obs}}$6, BFMT*'s bi-front expansion halves the wavefront radius per tree, yielding an exponential ($X_{\mathrm{obs}}$7) speed-up (Starek et al., 2015).

2. BiMotion as B-Spline Motion for Text-Guided Dynamic 3D Character Generation

BiMotion, as introduced in (Wang et al., 21 Feb 2026), addresses the generation of temporally coherent, expressive 3D motion sequences guided by natural language. Prior approaches used frame-discrete, fixed-length representations, limiting semantic richness and action diversity in generated motions.

BiMotion replaces discrete motion frames with a continuous, differentiable B-spline parameterization. A cubic B-spline curve $X_{\mathrm{obs}}$8 encodes variable-length $X_{\mathrm{obs}}$9-frame motions using a fixed number $$X_{\mathrm{free}} = \mathrm{cl}(X \setminus X_{\mathrm{obs}})$$0 of control points. Fitting $$X_{\mathrm{free}} = \mathrm{cl}(X \setminus X_{\mathrm{obs}})$$1 to control points $$X_{\mathrm{free}} = \mathrm{cl}(X \setminus X_{\mathrm{obs}})$$2 uses the closed-form Laplacian-regularized solver

$$X_{\mathrm{free}} = \mathrm{cl}(X \setminus X_{\mathrm{obs}})$$3

where $$X_{\mathrm{free}} = \mathrm{cl}(X \setminus X_{\mathrm{obs}})$$4 is the basis matrix, $$X_{\mathrm{free}} = \mathrm{cl}(X \setminus X_{\mathrm{obs}})$$5 is the discrete Laplacian, and $$X_{\mathrm{free}} = \mathrm{cl}(X \setminus X_{\mathrm{obs}})$$6 regularizes high-frequency variations.

The pipeline uses a B-spline VAE to reconstruct control points and a diffusion-Transformer to predict B-spline codes from text-encoded and geometry-encoded features. Quality is further enforced by a normal-fusion strategy, correspondence-aware loss, and a local-rigidity loss that stabilizes piecewise-geodesic lengths. The method is trained on BIMO, a dataset of 38,944 mesh-animation sequences (5–200 frames) with three text captions per sequence.

Evaluated on VBench, BiMotion attains best-in-class overall consistency, subject consistency, temporal flicker, and dynamic degree; requiring 4.4 s and 1.2 GB peak GPU memory, it outpaces AnimateAnyMesh (16.8 s, 3.1 GB) and other methods. In user studies, it achieves mean agreement, plausibility, and expressiveness scores of $$X_{\mathrm{free}} = \mathrm{cl}(X \setminus X_{\mathrm{obs}})$$7, $$X_{\mathrm{free}} = \mathrm{cl}(X \setminus X_{\mathrm{obs}})$$8, and $$X_{\mathrm{free}} = \mathrm{cl}(X \setminus X_{\mathrm{obs}})$$9 (5-point Likert), outperforming prior feed-forward and optimization-based baselines (Wang et al., 21 Feb 2026).

3. BiMotion in Bilingual Bi-Age-Group Speech Emotion Recognition

BiMotion in (Cahyawijaya et al., 2023) refers to a benchmark for evaluating the generalizability of speech emotion recognition (SER) models across language (English, Mandarin) and age (Adults: 20–59, Elderly: $\sigma : [0,1] \to X_{\mathrm{free}}$060). The benchmark aggregates six emotion-annotated corpora, partitioned by language and age-group. For English-Elderly, the training set includes data from CREMA-D, ElderReact, and TESS (total $\sigma : [0,1] \to X_{\mathrm{free}}$11,464 utterances).

Emotion recognition is posed as multi-label classification at the utterance level, aggregating original emotion schemes per underlying corpus (commonly basic emotions: anger, disgust, fear, happiness, sadness, neutral). Acoustic features consist solely of neural representations from XLSR-53 Wav2Vec 2.0, mean-pooled for utterance-level classification.

Two evaluation paradigms are defined: (1) Cross-group inference (“zero-shot”): train/test splits differ by age and/or language; (2) Cross-group data augmentation: augment in-group data with samples from other age/language sets and fine-tune accordingly. Weighted F1-score serves as the evaluation metric. In-group baselines yield average F1 $\sigma : [0,1] \to X_{\mathrm{free}}$2. Zero-shot cross-group transfer drops average F1 to $\sigma : [0,1] \to X_{\mathrm{free}}$3--$\sigma : [0,1] \to X_{\mathrm{free}}$4, while data augmentation (especially cross-age within-language) recovers or exceeds baseline performance (up to +5 pp on English).

Further, experiments on a new Cantonese-Adults benchmark (YueMotion) demonstrate that linguistic distance critically impacts cross-lingual SER transferability: Mandarin augmentation improves F1 from $\sigma : [0,1] \to X_{\mathrm{free}}$5, while English hurts ($\sigma : [0,1] \to X_{\mathrm{free}}$6). The absence of a harmonized annotation scheme or inter-annotator agreement is noted, as is the heterogeneity in speaker demographics and recording conditions across constituent datasets (Cahyawijaya et al., 2023).

4. Bilingual and Cross-Modal BiMotion in Text-to-Motion Synthesis

Bilingual text-to-motion generation further expands BiMotion’s referents, as in (Weng et al., 8 May 2025), where the problem is to synthesize 3D human motions aligned with bilingual (English, Chinese) textual input. The BiHumanML3D dataset, constructed by translating and validating HumanML3D captions, comprises 13,312 bilingual text–motion pairs.

The Bilingual Motion Diffusion model (BiMD) utilizes a continuous-time, score-based diffusion process conditioned on cross-lingually aligned text embeddings. Cross-lingual alignment between OpenCLIP (English) and XLM-Base (Chinese) is established by minimizing bidirectional KL divergence over embeddings. ReAlign, a reward-guided sampling method, addresses misalignment between text and motion by integrating a step-aware contrastive reward into the diffusion sampling SDE, thereby improving semantic consistency.

Empirical results on HumanML3D show BiMD+ReAlign achieves R@3 = 0.847, FID = 0.178, MM distance = 2.714, outperforming prior SOTA diffusion and latent consistency models. Ablations confirm the impact of reward-guided alignment and the importance of cross-lingual conditioning for Chinese text input. These results underline the pivotal role of the BiMotion problem in enabling semantically consistent, language-agnostic text-to-motion generation (Weng et al., 8 May 2025).

5. Comparative Methodologies and Empirical Outcomes

A tabular comparison across primary BiMotion usages:

Domain	Core Principle	Data/Representation	Evaluation/Performance
Sampling-based Planning	Bi-directional lazy FMT*	Random geometric graphs, cost-to-come/go	$\sigma : [0,1] \to X_{\mathrm{free}}$7 less time versus FMT*, $\sigma : [0,1] \to X_{\mathrm{free}}$8–$\sigma : [0,1] \to X_{\mathrm{free}}$9 faster than RRT*/PRM*, 100% success in OMPL tests (Starek et al., 2015)
Text-to-3D-Motion (Wang et al., 21 Feb 2026)	Continuous B-spline + diffusion	BIMO dataset, Laplacian B-spline, normal fusion	Best-in-class VBench metrics, user study scores $x_{\mathrm{init}}$0, 4.4 s GPU runtime (4$x_{\mathrm{init}}$1 faster than AnimateAnyMesh)
Speech Emotion Recognition (Cahyawijaya et al., 2023)	Bilingual, bi-age adaptation	Six-aggregate corpora, Wav2Vec 2.0	Cross-group inference F1$x_{\mathrm{init}}$240–50pp, augmentation F1$x_{\mathrm{init}}$35pp over baseline
Bilingual T2M (Weng et al., 8 May 2025)	Bilingual diffusion + reward alignment	BiHumanML3D, cross-lingual CLIP/XLM	R@3 0.84, FID 0.18 (BiMD+ReAlign); improved alignment in bilingual, cross-dataset settings

6. Impact, Limitations, and Research Trajectories

BiMotion, in the bi-directional or cross-modal sense, has advanced both algorithmic theory and practical performance for motion planning, 3D character generation, and robust speech emotion recognition across age and language boundaries. Notable limitations stem from reliance on aggregated or translated datasets (for emotion recognition), heterogeneous annotation protocols (no unified taxonomy or IAA), and inherent domain shifts in cross-lingual adaptation.

A plausible implication is that continued development of BiMotion approaches will focus on harmonizing taxonomies and annotation protocols, extending language and age coverage in benchmarks, and bridging domain gaps via adversarial or feature-level adaptation. In planning and motion synthesis, further sophistication of continuous trajectory representations and alignment losses is likely to yield richer, more semantically expressive models.

BiMotion thus represents an evolving, multifaceted paradigm for addressing bi-directional, bilingual, and cross-modal challenges in optimal control, animation, and perception systems, with verified improvements in efficiency, expressiveness, and generalizability as established in these primary sources.