Research map / Reinforcement learning / Imitation learning

Meta-Controller: Few-Shot Imitation of Unseen Embodiments and Tasks in Continuous Control

NeurIPS 2024 · Reinforcement learning › Imitation learning · approach: Behavior cloning theory and models · 2 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 Meta-Controller, a few-shot behavior cloning framework that uses joint-level tokenization, a structure-motion transformer encoder with embodiment/task-specific PEFT parameters, and a matching-based policy network to generalize to unseen embodiments and tasks from five reward-free demonstrations on DeepMind Control.

Problem

Few-shot behavior cloning that generalizes simultaneously to unseen robot embodiments and unseen tasks in continuous control, where prior modular policy methods lack few-shot task adaptation and few-shot IL methods handle only a single embodiment.

Approach

Tokenize states and actions per joint to unify heterogeneous embodiments, encode states with a structure encoder (joint axis) and a causal motion encoder (time axis) that split shared and embodiment/task-specific parameters via PEFT-style adapters, and predict actions with a non-parametric cross-attention matching against demonstration action features, fine-tuning only the adaptive parameters on n

Key results

• 5-shot average normalized score of 71.1 across eight held-out tasks, versus 37.9 for MetaMorph and 35.3 for MT-DT (best baselines).
• On the unseen reacher-four hard task, Meta-Controller scores 50.8 versus 8.5 for HDT and 3.8 for MetaMorph.
• Removing the structure encoder drops the 5-shot average from 71.1 to 31.6 in the architecture ablation.

Measured effects reported in the paper

Surprising findings

• Without adaptive embodiment-specific parameters in the structure encoder, the model still performs competitively on many tasks, suggesting joint-level I/O alone gives much compositional generalization.

Limitations

• Evaluated only in simulation (DeepMind Control) with 2D-plane embodiments; no real-world validation.
• Transformer encoders and decoders make inference costly for real-time resource-constrained robots.
• Struggles on tasks requiring reaching an unseen posture before reward appears, as shown in failure cases.

Experimental setup

Datasets: DeepMind Control suite replay-buffer demonstrations from DrQ-v2 expert agents

Benchmarks: DeepMind Control suite

Metrics: normalized score mean with standard error over 20 initial states

Baselines: FS-DT MT-DT PDT PDT+PEFT HDT L2M MetaMorph MTGv2

Builds on: Decision Transformer MetaMorph matching networks Hyper Decision Transformer Prompt Decision Transformer

Closest papers by meaning

• Zero-Shot Robotic Manipulation with Pre-Trained Image-Editing Diffusion Models (ICLR 2024 · similarity 0.88)
• Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (NeurIPS 2024 · similarity 0.88)
• OSVI-WM: One-Shot Visual Imitation for Unseen Tasks using World-Model-Guided Trajectory Generation (NeurIPS 2025 · similarity 0.87)
• Videos are Sample-Efficient Supervisions: Behavior Cloning from Videos via Latent Representations (NeurIPS 2025 · similarity 0.87)
• Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training (NeurIPS 2025 · similarity 0.87)
• What's the Move? Hybrid Imitation Learning via Salient Points (ICLR 2025 · similarity 0.87)