Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Online Continual Learning for Interactive Instruction Following Agents
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 Behavior-IL and Environment-IL continual learning setups for ALFRED-based instruction-following agents and a task-free Confidence-Aware Moving Average (CAMA) logit-update method for episodic-memory distillation.
Problem
Continual learning for embodied instruction-following agents that learn new behaviors and environments online, where stored distillation logits become outdated and task boundaries are unavailable.
Approach
Update episodic-memory logits via a moving average whose coefficients are set per class from recent ground-truth confidence scores of the current model, so no task boundary information is needed; the logits are then used for logit distillation.
Key results
- On ALFRED validation unseen, the proposed method reaches 14.19 SR_avg in Behavior-IL versus 12.04 for X-DER and 10.60 for DER++.
- In Environment-IL on validation unseen, the method reaches 15.67 SR_avg and 14.60 SR_last, versus 11.14 SR_avg for X-DER.
- Finetuning drops to 1.18 SR_last on unseen Behavior-IL, far below Joint training at 24.60.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use confidence-based dynamic coefficients to update stored logits rather than fixed coefficients (ablation) | CAM in Behavior-IL and Environment-IL · Validation seen and unseen | success rate and goal-condition success rate | Helped |
| Use confidence-based dynamic coefficients to update stored logits rather than fixed coefficients (ablation) | CAM · Environment-IL imbalanced validation unseen | SR_avg: 16.02 ± 1.20 -> 17.24 ± 1.78 | Helped |
| Update stored logits using confidence-aware moving averages rather than leaving logits unupdated | DER++ comparison in Behavior-IL and Environment-IL · Seen and unseen validation | success rate and goal-condition success rate | Helped |
| Use confidence-aware logit updates rather than X-DER's updates limited to novel classes | CAM versus X-DER · Behavior-IL and Environment-IL | success rate and goal-condition success rate | Helped |
Surprising findings
- On validation seen, the variant without dynamic coefficients slightly beats the full method on some metrics (e.g., Behavior-IL SR_avg 29.89 vs 29.67), and the ablation shows the gain is not uniform.
- Standard replay methods such as MIR can match or beat distillation baselines like DER++ and X-DER on some seen metrics.
Limitations
- Only class prediction is continually updated; object mask generation is assumed available and left for future work.
- Evaluated only on disjoint task sequences; overlapping (blurry) task streams are not studied.
- Absolute success rates remain low on unseen environments (about 14-16% SR), well below Joint training.
Experimental setup
Datasets: ALFRED
Benchmarks: ALFRED validation seen ALFRED validation unseen
Metrics: SR GC SR_last SR_avg GC_last GC_avg
Baselines: Finetuning EWC++ ER MIR CLIB DER++ X-DER Joint
Builds on: DER++ X-DER ALFRED ER MIR
Closest papers by meaning
- CCIL: Continuity-Based Data Augmentation for Corrective Imitation Learning (ICLR 2024 · similarity 0.87)
- Autonomous Continual Learning for Environment Adaptation of Computer-Use Agents (NeurIPS 2026 · similarity 0.86)
- NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks In Open Domains (ICLR 2025 · similarity 0.86)
- Continual Reinforcement Learning by Planning with Online World Models (ICML 2025 · similarity 0.85)
- Robot-R1: Reinforcement Learning for Enhanced Embodied Reasoning in Robotics (NeurIPS 2025 · similarity 0.85)
- Behavioral Exploration: Learning to Explore via In-Context Adaptation (ICML 2025 · similarity 0.85)
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.