Reproduce “Mind Dreamer: Untethering Imagination via Active Causal Intervention on Latent Manifolds” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shaojun Xu, Xiaoling Zhou, Yihan Lin, Yapeng Meng, Xinglong Ji, Luping Shi, Rong Zhao
Mind Dreamer untethers imagination in Model-Based Reinforcement Learning by enabling active causal intervention on latent manifolds, allowing agents to proactively explore unmastered state regions. This approach achieved an average 1.67x speedup in sample efficiency on DeepMind Control Suite tasks compared to DreamerV3, with up to 8.8x acceleration in challenging sparse-reward environments, enhancing overall performance.
Key insights
- Untethering imagination through ACI allows agents to proactively explore unmastered regions of the latent manifold, accelerating discovery and policy learning.
- The combination of pragmatic (Relay Value) and epistemic (Relay Uncertainty) potentials, combined with structural manifold constraints, ensures targeted, physically plausible, and efficient exploration.
- A quadratic discount factor (γ²) for uncertainty propagation is necessary to establish a stable "Epistemic Horizon," preventing the generator from chasing distal, unverified information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.