Unified Prediction and Planning via Conflict-Aware Disjoint Parameter Training

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Expert, extended

Summary

A novel framework, Disjoint Parameter Training (DPT) combined with Sparse Merging, addresses "Skill Conflict" in compact, unified prediction and planning models for social robot navigation. Skill Conflict arises when distinct tasks, like predicting neighbor behaviors and ego-centric safety planning, compete for shared parameters in resource-constrained edge devices. DPT mitigates this by separating key parameter regions for each task during fine-tuning using binary masks, ensuring their core capabilities are preserved. Sparse Merging then selectively integrates only the most influential parameters, preventing interference. Evaluated on JRDB and JTA crowd navigation benchmarks, the framework demonstrates superior performance over baselines such as DIPP and DTPP. It also shows broad applicability to other merging methods and extends to end-to-end autonomous driving, achieving an inference latency of 10.4 ms with 6.44M parameters on an NVIDIA RTX 3090.

Key takeaway

For Machine Learning Engineers deploying unified prediction and planning models on resource-constrained edge devices, you should adopt Disjoint Parameter Training (DPT) and Sparse Merging. This approach explicitly mitigates "Skill Conflict" by separating task-specific parameter regions during training and selectively merging only core components, ensuring both accurate prediction and safe planning. Consider integrating a joint reasoning stage for further performance gains.

Key insights

Compact unified models for robot navigation suffer "Skill Conflict" from shared parameters, resolved by DPT and Sparse Merging.

Principles

• Task-specific parameters conflict in compact models.
• Disjoint training prevents parameter overlap.
• Sparse merging enhances multi-task composability.

Method

Disjoint Parameter Training (DPT) fine-tunes task-specific models by distributing parameter learning regions with binary masks. Sparse Merging then integrates only the most salient task-vector components.

In practice

• Apply DPT to separate prediction and planning parameters.
• Use sparse merging for efficient model unification.
• Integrate joint reasoning for enhanced performance.

Topics

• Robot Navigation
• Motion Prediction
• Motion Planning
• Model Merging
• Disjoint Parameter Training
• Skill Conflict
• Edge Devices

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Robotics Engineer, Machine Learning Engineer

Related on AIssential

• Unified Prediction and Planning via Conflict-Aware Disjoint Parameter Training — DPT resolves "Skill Conflict" in unified robot navigation models by separating task-specific parameters for improved performance.
• Diverse Yet Consistent: Context-Guided Diffusion with Energy-Based Joint Refinement for Multi-Agent Motion Prediction — Context-guided diffusion with energy-based refinement improves diverse yet consistent multi-agent motion prediction.
• A Mechanistic Analysis of Sim-and-Real Co-Training in Generative Robot Policies — Co-training effectiveness in robotics stems from structured representation alignment and importance reweighting.
• CLASP: Language-Driven Robot Skill Selection and Composition using Task-Parameterized Learning — CLASP integrates TP-KMPs with VLMs for data-efficient, language-driven robot skill learning, selection, and composition.
• Enough is as good as a feast: A Comprehensive Analysis of How Reinforcement Learning Mitigates Task Conflicts in LLMs — Reinforcement Learning (RL) training inherently reduces task conflicts in LLM model merging, leading to better performance preservation.
• When Model Merging Breaks Routing: Training-Free Calibration for MoE — MoE model merging often fails due to routing breakdown; HARC offers a training-free, curvature-aware calibration to fix it.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.