AFTER: A World Model with Adaptive Fast--Slow Timescales for Error-Driven Retention
Organizations: Tongji University, Shanghai, China
Abstract
In robotics and embodied AI, a core challenge for world models is adapting to sudden, persistent, yet initially unknown changes in robot dynamics during interaction. For instance, when a robotic arm grasps an object with an unknown payload, the robot dynamics change suddenly. If the world model fails to adapt, its prediction errors accumulate over the open-loop rollout and ultimately mislead the robot's actions. Existing world models, however, struggle to simultaneously satisfy three critical requirements for dynamics adaptation: responsiveness, reacting promptly to sudden changes to curb prediction error accumulation; stability, distinguishing persistent changes from transient disturbances to prevent incorrect updates; and persistence, retaining the adapted dynamics without new observations so that later predictions do not revert to the old dynamics. Inspired by predictive coding in the human brain with multi-timescale adaptation, we propose AFTER-WM, a world model with Adaptive Fast-slow Timescales for Error-driven Retention, whose prediction errors drive a fast correction that fades and a slow correction that persists, while dynamically adjusting the combination weights of pre-trained experts without altering them, to overcome the trade-offs among responsiveness, stability and persistence. On five simulated robot datasets and real UAV data, AFTER-WM has a lower multi-step mean error than all ten compared models and the highest mean return in a joint-failure control task. In simulation, its multistep mean error is 16%-34% below the strongest compared model's and is also the lowest 1-4 steps after a change and under transient disturbances alone, and its step-40 error is the lowest as well, reflecting responsiveness, stability and persistence. All code (method, baselines, experiments) and datasets will be released online.
Figures & tables
| Horizon: 40 steps | JSHC | ESCP-HC | Panda | UR5e | Hopper | |||||
| Model | multi-step mean | step 40 | multi-step mean | step 40 | multi-step mean | step 40 | multi-step mean | step 40 | multi-step mean | step 40 |
| DreamerV3-style RSSM | 0.197 0.003 | 0.311 0.004 | 0.372 0.006 | 0.532 0.009 | 0.329 0.016 | 0.404 0.022 | 0.331 0.011 | 0.418 0.013 | 0.145 0.003 | 0.224 0.009 |
| TD-MPC2 | 0.192 0.002 | 0.370 0.012 | 0.284 0.008 | 0.599 0.033 | 0.189 0.013 | 0.411 0.026 | 0.226 0.014 | 0.547 0.064 | 0.102 0.004 | 0.204 0.010 |
| TD-MPC2-40 | 0.212 0.002 | 0.279 0.009 | 0.332 0.002 | 0.499 0.012 | 0.244 0.007 | 0.409 0.010 | 0.253 0.010 | 0.450 0.022 | 0.111 0.002 | 0.219 0.013 |
| MLP | 0.195 0.001 | 0.306 0.009 | 0.279 0.002 | 0.486 0.009 | 0.0841 0.0024 | 0.130 0.007 | 0.0846 0.0035 | 0.116 0.003 | 0.0953 0.0011 | 0.177 0.004 |
| GrBAL | 0.210 0.006 | 0.402 0.021 | 0.352 0.087 | 2.119 3.354 | 0.0889 0.0040 | 0.176 0.009 | 0.0836 0.0034 | 0.145 0.006 | 0.0817 0.0040 | 0.203 0.032 |
| Setting | Compared models beaten | Smallest margin (closest model) |
| Beyond training (ESCP-HC) | 10 of 10 | 5.9% (ReBAL) |
| No change (5 datasets) | 10 of 10 | 6.6% (ReBAL) |
| Real flights (ALFA) | 10 of 10 | 13.1% (MOLe) |
| Control (JSHC) | 10 of 10 | +13.8 (MOLe) |
Appendix figures & tables46 assets
Supplementary material from the paper’s appendix.
Appendix
| JSHC | ESCP-HC | Panda | UR5e | Hopper | ALFA | |
| Source | joint-disable protocol following GrBAL [ 7 ] and MBCD [ 1 ] | sudden-change protocol following ESCP [ 6 ] | self-built | self-built | self-built | real recorded flights [ 5 ] |
| Robot model | HalfCheetah (Gymnasium HalfCheetah-v5, MuJoCo 3.11) | HalfCheetah (Gymnasium HalfCheetah-v5, MuJoCo 3.11) | Franka Panda (MuJoCo Menagerie), payload body at the tool point | UR5e (MuJoCo Menagerie), payload body at the flange | Hopper (Gymnasium Hopper-v5 model), joint position servos | CMU fixed-wing UAV |
| State | 17: torso height, pitch, 6 joint angles, 9 velocities | 17: as JSHC | 14: 7 joint angles, 7 joint velocities | 12: 6 joint angles, 6 joint velocities | 11: torso height, pitch, 3 joint angles, 6 velocities | 12: attitude (2), heading sin/cos, body rates (3), NED velocity (3), altitude, airspeed |
| Action | 6: joint torques | 6: joint torques | 7: compliant joint-servo targets | 6: compliant joint-servo targets | 3: joint-servo targets | 3: autopilot roll, pitch, airspeed commands |
| Change channels | one of 5 joints has zero torque | torso mass, joint damping, gravity, each | payload 0 or 0.5–2.5 kg; joint friction 0.1–3; motor strength 1 or 0.8–1.2 | payload 0 or 1–4 kg; joint friction 0.1–3; motor strength 1 or 0.8–1.2 | ground friction 0.1–3; servo (motor) strength 1 or 0.6–1.4; torso mass 1 or 1.3–2.5 | engine or control-surface failure at a labeled time; some flights without failure |
| Switch schedule | failed joint re-drawn every 100 steps, random phase | new task every 100 steps (ESCP schedule) | 6 sequence types on segments [0,40,120,190,260,300], incl. drift | as Panda | as Panda | failure persists after onset |
| Model | Main sizes (range over datasets) | Optimizer (native LR) | LR mult. | Peak LR | Clip | Chosen per dataset on validation |
| AFTER-WM (ours) | Base (fixed after stage 1): router , 3 experts (434–447k). Online: update networks and , each ; fast readout per expert; fast state 32 and slow state 3 per expert; hidden widths of the measurement network 86–108 and 116–146, of the slow readout 73–92 and 75–94, and of the filter network 18–23 | Adam ( ) | 1.0 | inverse routing temperature per seed: JSHC 0.127–0.15, ESCP-HC 0.15–0.2, Panda 0.1, UR5e 0.05–0.15, Hopper 0.2–0.3, ALFA 0.127–0.2 | ||
| TD-MPC2 | latent 128 (SimNorm 8), encoder , dynamics/decoder width 480–482, consistency horizon 3 | Adam ( ; encoder ) | 20 | – | ||
| TD-MPC2-40 | as TD-MPC2; consistency/decoder horizon 40 | Adam ( ; encoder ) | 20 | – | ||
| ReBAL | residual MLP recurrent cell, width 580–584 | Adam ( ) | 0.1 | 1.0 ∗ | – | |
| GrBAL | MLP 581–585; one inner step on the last 16 transitions | Adam, outer ( ) | 1.0 ∗ | inner step: JSHC 0.1, ESCP-HC , Panda 0.1, UR5e 0.01, Hopper 0.1, ALFA 0.01 | ||
| MOLe | MLP 581–585; inner step 0.01, 0.01, CRP 1 | Adam, outer ( ) | 1.0 ∗ | – |
World Model Learning
World model learning builds predictive models of how environments change under actions to support planning, simulation, and decision-making.
See all 456 papers