AFTER: A World Model with Adaptive Fast--Slow Timescales for Error-Driven Retention
This preprint introduces AFTER-WM, a world model that adapts to changing robot dynamics by combining fading fast corrections, persistent slow corrections and error-based expert weighting. Across five simulated datasets, its multi-step prediction error was 16.4% to 33.7% below the strongest baseline, suggesting a way to improve planning after payload changes or actuator faults.
Can a world model respond quickly to persistent dynamics changes, reject transient disturbances and retain its adaptation throughout predictions without further observations?
An unfamiliar payload, changing friction or a failed actuator can invalidate a robot's learned dynamics. Prediction errors then compound as a planner rolls the model forward without observations. Aggressive adaptation can mistake a brief impact for a lasting change, while corrections that fade quickly can lose useful adaptation before the prediction horizon ends.
TD-MPC2 lacks a dedicated mechanism for identifying dynamics changes, while DreamerV3-style RSSM updates a recurrent state without explicit dynamics adaptation. GrBAL and MOLe adapt model parameters, and ReBAL uses recent history, but their single adaptation timescale creates a responsiveness versus stability trade-off. MTS3 separates fast and slow variables, but updates its slow variable at fixed intervals, delaying its response.
AFTER-WM combines 3 pre-trained MLP dynamics experts with error-driven memories. Training first learns the experts and router, then freezes them while training correction and memory networks. During inference, observed prediction errors update expert weights, a fading fast state and an uncertainty-sensitive slow state without gradient updates. Each prediction resets these memories and reconstructs them from 64 context steps. During the 40-step horizon (2 s), expert weights and slow states remain fixed while fast states decay. Neural models in the comparison have about 0.70M parameters.
Against ten baselines, AFTER-WM led on all five MuJoCo datasets over 5 training seeds. Multi-step mean nMSE versus each dataset's strongest baseline was: JSHC, 0.124 vs ReBAL 0.166, 25.7% lower; ESCP-HC, 0.208 vs ReBAL 0.249, 16.4% lower; Panda, 0.0473 vs MOLe 0.0711, 33.5% lower; UR5e, 0.0435 vs ReBAL 0.0641, 32.1% lower; Hopper, 0.0499 vs MOLe 0.0752, 33.7% lower. This metric averages cumulative mean errors at horizons 1, 5, 10, 20 and 40, using training-standardized states. Step-40 error was 21.9% to 41.9% below those same baselines. On recorded real ALFA UAV flights, prediction error was 13.1% below the closest baseline, MOLe, across 4 seeds. In simulated JSHC joint-failure control, evaluated with 5 seeds and 48 test episodes using the same cross-entropy-method planner, mean return reached 93.1, exceeding MOLe by 13.8.
Closed-loop control was tested only in simulation on one joint-failure task. Real-world evidence consists of prediction on recorded fixed-wing UAV flights, without live control experiments. Inputs are numerical states, not images; image observations and real-robot deployment are future work. Memories are rebuilt from a finite context for each prediction, so retention is demonstrated within the rollout rather than as indefinite online learning. Code and dataset release is promised, but availability is not confirmed in the paper. Inference latency and compute cost: Not reported in the paper.
Robot-arm integrators, adaptive-control suppliers and UAV autonomy developers could now evaluate this approach in research planning systems that already expose state estimates and action histories. Product use would require live hardware control validation, reliable state observations and acceptable runtime cost. A commercial deployment timeline is not established. Not reported in the paper.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.