LoopWAM: Closed-Loop Generalist Policy Improvement from Deployment Rollouts
We introduce LoopWAM, a closed-loop world-action model that learns directly from this full quality spectrum, using rollouts from both itself and other deployed policies.
ProofPaper ↗
Key points
- Scaling expert demonstrations alone does not enable generalist robot policies to improve from deployment.
- Real-world rollouts contain successes, suboptimal behavior, and failures: imitating every recorded action can reinforce poor behavior, while filtering out failures discards evidence of physical interactions and their consequences.
- Source-dependent noise sampling unifies behavior learning and action-conditioned world modeling, while quality steering and quality-weighted supervision retain full video supervision across qualities and regulate action imitation.
- LoopWAM also achieves 94.5% average success on RoboTwin 2.0 and 49.5% overall success on RoboCasa365 without additional embodied pretraining, obtaining the best overall results among the compared methods across all three benchmarks.
Sources (1)
- [1]LoopWAM: Closed-Loop Generalist Policy Improvement from Deployment RolloutsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 05:59 PM
We introduce LoopWAM, a closed-loop world-action model that learns directly from this full quality spectrum, using rollouts from both itself and other deployed policies.
Scaling expert demonstrations alone does not enable generalist robot policies to improve from deployment.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026Run fast, private agents on-device with LFM2.5-24B-A2B
- Oct 13, 2026googleapis/python-genai v2.30.0
- Oct 13, 2026Hybrid Reasoning Agents
- Oct 13, 2026Fish Audio Drama 3 Technical Report
- Oct 13, 2026WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction
- Oct 13, 2026WorldRecon: Generative Multi-View Mesh Reconstruction for Large-Scale 3D World