Yurui Du

Yurui Du, Pinhao Song, Yutong Hu et al.

A central challenge of visual control with model-based reinforcement learning (RL) is reliable long-horizon planning: long rollouts with learned latent dynamics exhibit branching futures and multi-modal action-value distributions. In addition, compounding model errors amplified by visual occlusions make deep imagination brittle. We present ELVIS, a latent model predictive controller (MPC) designed to make long-horizon planning practical. ELVIS plans in a Dreamer-style recurrent state space model (RSSM) and replaces standard unimodal model predictive path integral (MPPI) with a Gaussian-mixture MPPI that maintains multiple coherent hypotheses over long horizons, avoiding mode averaging under branching rollouts. In parallel, ELVIS stabilizes deep imagination with a shared uncertainty-aware lambda-return: an ensemble of latent critics defines an upper-confidence-bound (UCB) score that gates a time-varying lambda, adaptively trading off bootstrapping versus look-ahead to limit compounding error during planning. The same return is used both to train an actor-critic prior from imagined rollouts and to score candidate trajectories inside GMM-MPPI, aligning RL objectives with the planner's long-horizon optimization. On fourteen DeepMind Control Suite visual tasks, ELVIS establishes state-of-the-art performance compared with TD-MPC2 and DreamerV3. Finally, ELVIS transfers zero-shot to a real-world sand-spraying task with severe occlusions, improving surface-quality metrics and demonstrating robustness beyond simulation.

Yurui Du, Flavia Sofia Acerbo, Jens Kober et al.

In recent years, imitation learning (IL) has been widely used in industry as the core of autonomous vehicle (AV) planning modules. However, previous IL works show sample inefficiency and low generalisation in safety-critical scenarios, on which they are rarely tested. As a result, IL planners can reach a performance plateau where adding more training data ceases to improve the learnt policy. First, our work presents an IL model using the spline coefficient parameterisation and offline expert queries to enhance safety and training efficiency. Then, we expose the weakness of the learnt IL policy by synthetically generating critical scenarios through optimisation of parameters of the driver's risk field (DRF), a parametric human driving behaviour model implemented in a multi-agent traffic simulator based on the Lyft Prediction Dataset. To continuously improve the learnt policy, we retrain the IL model with augmented data. Thanks to the expressivity and interpretability of the DRF, the desired driving behaviours can be encoded and aggregated to the original training data. Our work constitutes a full development cycle that can efficiently and continuously improve the learnt IL policies in closed-loop. Finally, we show that our IL planner developed with less training resource still has superior performance compared to the previous state-of-the-art.