Research map / Probabilistic and scientific modeling / Dynamical systems
Model Adaptation for Time Constrained Embodied Control
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 MoDeC, a time-constraint-aware modular network for multi-task embodied control that selects modules iteratively under device-specific latency constraints, distills the selector into a single step, and adapts to devices with a few-shot device adapter.
Problem
Embodied control agents running multiple tasks must meet per-task inference latency constraints on heterogeneous embedded devices, but dynamic/adaptive model structures have not been studied for this setting.
Approach
Formulate adaptation to time and resource constraints as dynamic module routing in a soft-modularization multi-task network: an iterative module selector (trained with Reptile+REINFORCE and an action-similarity reward) is distilled into a single-step selector, and a few-shot device adapter maps latency constraints to module counts.
Key results
- On CARLA, MoDeC achieves 14.4% higher success rate than DS-Net while keeping constraint violations under 1%.
- On Meta-world single-task, MoDeC gains 25.1% over the strongest baseline D2NN across device/constraint settings.
- On AI2THOR unseen configurations, MoDeC outperforms DS-Net by a gap of 15.4% to 70.8%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MoDeC instead of DS-Net | Autonomous driving control · CARLA | success rate: 14.4% | Helped |
| Use MoDeC instead of D2NN | Multi-task control · Meta-world MT10 | success rate: 5.7% | Helped |
| Use MoDeC instead of DS-Net | Object navigation · AI2THOR Unseen configurations | success rate: 15.4% to 70.8% | Helped |
| Remove distillation and use only iterative module selection (ablation) | MoDeC · Meta-world | success rate | Hurt |
| Remove distillation and use only iterative module selection (ablation) | MoDeC · Meta-world | inference time | Hurt |
| Learn single-step module selection directly without distillation (ablation) | MoDeC · Meta-world | success rate | Hurt |
| Use a 2 × 8 base network instead of 4 × 4 (ablation) | MoDeC · Meta-world | success rate | Hurt |
| Use an 8 × 2 base network instead of 4 × 4 (ablation) | MoDeC · Meta-world | success rate | Hurt |
Surprising findings
- Reward including environment reward degraded the iterative module selector, so the selector was trained with action-similarity reward only.
Limitations
- Baselines DRNet and D2NN require separate networks per constraint, so comparison setup differs from the single-model MoDeC.
- Future work noted on learning complex constraints from instructions, including safety and resource limits.
Experimental setup
Datasets: Meta-world MT10 CARLA (12 maps) AI2THOR rearrangement
Benchmarks: Meta-world MT10 CARLA AI2THOR
Metrics: success rate FLOPs inference time constraint violation rate
Baselines: DRNet D2NN DS-Net RL-AA
Builds on: soft modularization Reptile REINFORCE SAC DS-Net
Closest papers by meaning
- Robot-R1: Reinforcement Learning for Enhanced Embodied Reasoning in Robotics (NeurIPS 2025 · similarity 0.86)
- VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.85)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.85)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.85)
- Meta-Controller: Few-Shot Imitation of Unseen Embodiments and Tasks in Continuous Control (NeurIPS 2024 · similarity 0.85)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.84)
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.