Multimodal Contrastive Learning for Zero-Shot Instruction-Following Robot with Synthetic Data

Received: 20 February 2026 | Revised: 22 March 2026 and 1 April 2026 | Accepted: 3 April 2026 | Online: 26 April 2026

Corresponding author: Washington Kamadi

Abstract

Robot trajectory prediction is heavily dependent on large-scale real-world demonstrations, which limit scalability, increase data acquisition costs, and eventually prevent zero-shot generalization. To address this limitation, this paper introduces Zero-Shot Task Learning (ZSTL), a multimodal framework that uses structurally aligned synthetic data and contrastive learning to enable instruction-based trajectory generation without reliance on real-world demonstrations. ZSTL jointly encodes natural-language instructions, depth observations, LiDAR-derived spatial representations, and action trajectories within a joined embedding space, allowing cross-modal alignment and conditional behavior synthesis. The proposed architecture preserves modality structure prior to fusion by representing depth inputs as spatial tokens and LiDAR observations as temporal tokens. Together with a text token, these form a 101-token multimodal context attended over by a Transformer decoder to predict full 50-step trajectories with Gaussian uncertainty estimates. The system integrates a pretrained Bidirectional Encoder Representations from Transformers (BERT) language encoder, a ResNet-18 depth backbone, a one-dimensional convolutional LiDAR sequence encoder, and a two-layer Transformer decoder comprising approximately 125M parameters. Training was conducted entirely on a procedurally generated synthetic dataset of 5,000 samples for 50 epochs. The results demonstrate stable convergence, with the trajectory-negative log likelihood decreasing from 3.465 to -0.695 on validation data and the combined loss reaching -0.540 at epoch 18 under cosine annealed learning. The contrastive objective (InfoNCE, τ = 0.07) stabilized near 1.55, indicating consistent cross-modal alignment. The trajectory evaluation yielded an average final position error of 9.97 cm, a collision-free execution rate of 65.9%, and a task success rate of 59.2%, showing that structured synthetic supervision can support physically meaningful motion generation.