World models learned from offline data allow an agent to plan without further interaction with the environment, but the sampling-based planners used with them handle safety constraints only by scoring or rejecting sampled action sequences. We instead pose safe planning in the latent space of a learned world model as a successive convex program (SCP), in which a learned discrete-time control barrier function is imposed as a constraint at every step of the horizon. SCP requires smooth dynamics, so we build the networks from smoothed MaxMin activations and spectrally normalised layers, and bound the resulting linearisation error in terms of the Lipschitz bound and the smoothing width. As SCP is a local method, we start it from a sampled plan. On two Safety-Gymnasium navigation tasks with offline data, this planner has fewer violations than MPPI and CEM drawing 128 times as many samples at equal wall-clock time in our main comparison, though not significantly, and earns greater return at longer horizons, and is much safer than SCP without a sampled start. SCP converged more often on Lipschitz-bounded models. Unlike sampling, it also reports at each step whether its plan is a feasible, first-order optimal KKT point of the planning problem in the model, or that no feasible plan was found nearby. Our implementation is publicly available at https://github.com/youngers2006/scwm.
Figures & tables
Baseline
Pairs
Δ violations (pp)
P (safer)
Δ return
P (higher return)
Main
SCP
240
−36.7 [ −43.8 , −28.8 ]
>0.999
+0.17 [ −0.05 , +0.40 ]
0.94
MPPI
240
−3.8 [ −9.0 , +1.5 ]
0.92
+0.07 [ −0.13 , +0.25 ]
0.78
CEM
240
−3.3 [ −8.4 , +1.7 ]
0.90
+0.18 [ +0.06 , +0.31 ]
0.999
N=10
SCP
30
−40.0 [ −57.9 , −16.1 ]
0.999
+0.49 [ −0.05 , +1.02 ]
0.96
MPPI
30
−16.7 [ −29.7 , −3.2 ]
0.98
−0.10 [ −0.61 , +0.33 ]
0.36
CEM
30
−10.0 [ −21.4 , +1.0 ]
0.94
+0.21 [ −0.17 , +0.57 ]
0.88
Table 1: SCP+MPPI minus SCP and each time-matched sampler. Main: the main comparison ( c=1.5 , N=10 , both tasks, ϵ≥0.1 ). N=10 , 20, 30: PointGoal1 at each horizon ( c=1.5 , ϵ=0.2 ; 30 pairs), where at N=20 and 30 there are too few violations to compare safety. Negative Δ violations and positive Δ return favour SCP+MPPI; brackets are 95% intervals.
Figure 1: Violation rate (left pair) and return (right pair) against the margin ϵ on each task’s c=1.5 model at N=10 , with 95% intervals.
Figure 2: SCP+MPPI’s convergence at N=10 . Left pair: how its steps ended on each task’s c=1.5 model (certified: converged with the plan’s rollout meeting the barrier). Right pair: the share of its steps at the 40-iteration cap for each Lipschitz bound.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: The two tasks from above, one layout each (layouts are drawn from the episode’s seed). The point robot (red) must reach the goal (green) without entering a hazard (blue). In PointGoal2, touching a vase (cyan) is also a violation.
Latent dimension nz
32
Encoder input
53 of the 60 observation channels (drops the accelerometer and the out-of-plane velocity, angular-rate and magnetometer components)
Dynamics, reward, value, barrier
MLP, 2×256 , smoothed MaxMin ( δ=0.05 ), the network bound c is split evenly over the three layers ( c1/3 each), the dynamics network is residual, z+gψ(z,u) with gψ bounded by c−1
QΘ (training only)
MLP, 2×256 , unbounded
Reward target
dense part of the reward, r−\mathds1[r>0.5] , the goal bonus is left to the value
Table 5: As Table 1 , on the two unbounded ( c=∞ ) models.
Figure 4: Fig. 1 with the samplers at both budgets: 512 sequences per round (open markers, dashed) and time-matched (filled, solid). Violation rate (top) and return (bottom) against the margin ϵ on each task’s c=1.5 model at N=10 , with 95% intervals.
Figure 5: Return against violation rate on each task’s c=1.5 model at N=10 , pooled over ϵ≥0.1 , with 95% intervals. Arrows lead from each sampler with 512 sequences per round to the same sampler time-matched.
Figure 6: Violation rate against the margin ϵ for every planner on every N=10 world model (rows: tasks; columns: Lipschitz bound c ), with 95% Wilson intervals.
Figure 7: Return against the margin ϵ for every planner on every N=10 world model (rows: tasks; columns: Lipschitz bound c ), with 95% bootstrap intervals.
Figure 8: Violation rate (left) and return (right) against the horizon N on PointGoal1 ( c=1.5 , ϵ=0.2 ), with the samplers at 512 sequences per round (open markers, dashed) and at their time-matched budget (filled, solid), with 95% intervals. Each N uses its own world model, trained on windows of N+1 steps, 30 episodes per point.
Figure 9: Return against mean plan time per step at N=20 and 30, with 95% intervals. Each sampler is shown at 512 sequences per round and at its time-matched budget (along the bottom), SCP+MPPI and SCP are single points.
Figure 10: Share of steps followed by a violation within N steps, after a plan that met the barrier in the model (dots) and after one that did not (crosses), pooled over the six N=10 models and ϵ≥0.1 , with 95% Wilson intervals (log scale).
Figure 11: Violations per 100 episodes over the six N=10 models and ϵ≥0.1 (720 episodes per planner), split into those preceded within N steps by a plan that failed the barrier in the model (warned, light) and those that were not (not warned, dark, counts above the bars).
Figure 12: Share of episodes without a violation by each step, at each margin, on PointGoal1’s c=1.5 model at N=10 . Episodes end early only at a violation.
Figure 13: Model optimism: each executed plan’s predicted return minus the discounted return received from that step on, averaged over each episode, with 95% bootstrap intervals. N=10 pools every world model ( ϵ≥0.1 ); N=20 and 30 are the horizon study’s models.
Safe motion planning develops methods for robots and vehicles to reach goals while avoiding collisions and respecting dynamic and uncertainty constraints.
Lyles School of Civil and Construction Engineering, Purdue University, West Lafayette 47907, USA · Department of Civil and Environmental Engineering, University of Wisconsin–Madison, Madison 53706, USA