Quadruped parkour policies handle canonical obstacles well, but their training curricula often miss the irregular, compound geometry of industrial and disaster-response sites, where agile locomotion matters most. Such environments pose a hard generalization problem: even training directly on reconstructed geometry from the target site does not reliably yield policies that traverse it. Hand-designed curricula span only a few terrain families, and recent LLM-generated curricula offer little control over which policy weaknesses they target. We propose FCTS, a failure-conditioned terrain curriculum framework that uses localized traversal failures to automatically generate the terrains a policy should train on next. Each failure's location, termination mode, robot state, and surrounding terrain geometry condition this generation, and Monte Carlo Tree Search explores alternative curriculum trajectories under a fixed training budget. On terrains built from scanned real-world construction sites, FCTS raises success rate from 28.4% to 85.8% on a medium-difficulty course and from 14.9% to 44.7% on a hard course relative to the strongest parkour or LLM-curriculum baselines, and maintains higher progress on extreme terrains where full traversal is rare. Hardware experiments on a quadruped robot confirm that the learned FCTS policies transfer to the real world.
Figures & tables
Fig. 1 : FCTS traverses reconstructed real-world terrain on which baseline curriculum stop early. The site is an unseen deployment environment, held out from curriculum generation and training, and all three policies start from the same pre-trained checkpoint under the same commanded route, shown as overlaid poses through waypoints marked in red. Each trajectory is a representative episode.
Fig. 2 : Overview of our failure-conditioned curriculum search (FCTS). Validation failures guide LLM terrain generation, while parallel MCTS explores alternative policy-improvement trajectories.
Fig. 3 : Evaluation routes in the two reconstructed environments. Left (Debris site) shows both the relief-stratified traversal routes (Courses 1–8) and directly specified traversal routes (Courses 9–11) across all difficulty levels. Right (Exposed Beam site) shows the directly specified evaluation routes only, since the relief-based course generator is not reliable on its large gaps and disconnected structures.
Difficulty
Route (m)
Mean Relief (m)
P95 Relief (m)
Max Step (m)
Easy
11.4–20.2
0.128–0.144
0.275–0.295
0.26–0.32
Mid
10.4–15.4
0.204–0.210
0.343–0.389
0.32–0.43
Hard
9.6–19.2
0.290–0.303
0.459–0.503
0.58–0.76
V-Hard
13.7–17.8
0.318–0.325
0.589–0.629
0.51–0.87
TABLE I : Geometry of the relief-stratified routes in the Debris site benchmark . Values show the range across the two courses in each difficulty level. Relief denotes local terrain height variation within a 0.4m robot footprint; P95 captures the upper tail of terrain difficulty.
Difficulty
Method
Success (%) ↑
Progress IQM ↑
WP (%) ↑
Easy
Extreme Parkour
29.8
0.77
74.1
Eurekaverse
29.4
0.45
26.8
Modified Eurekaverse
31.5
0.61
55.4
FCTS (Ours)
84.0
1.00
100.0
Mid
Extreme Parkour
23.9
0.63
53.2
Eurekaverse
28.4
0.70
67.5
TABLE II : Performance on relief-stratified routes in the Debris site , with two routes per difficulty tier. Each route is evaluated over two independent seeds with 256 trials per seed. WP denotes median waypoint completion normalized by the number of route waypoints. Higher is better ( ↑ ).
Fig. 4 : Failure cases and the corresponding winning FCTS curriculum branch. Top: three terrain regions where Extreme Parkour and Eurekaverse fail while FCTS completes the traversal. Bottom: the highest-valued branch found by FCTS and the terrain families generated along it. The branch progressively introduces irregular rubble, inclined support, repeated height-and-gap transitions, and their composition.
Difficulty
Method
Success (%) ↑
Progress IQM ↑
WP (%) ↑
Debris site
Easy
Extreme Parkour
0.8
0.27
30.0
Eurekaverse
2.3
0.31
40.0
Modified Eurekaverse
1.2
0.21
20.0
FCTS (Ours)
36.7
0.59
50.0
Mid
Extreme Parkour
5.1
0.16
0.0
TABLE III : Performance on directly specified routes in the Debris site and Exposed Beam Site . Each route is evaluated over two independent seeds with 256 trials per seed. WP denotes median waypoint completion normalized by the number of route waypoints. Higher is better ( ↑ ).
Fig. 5 : Real-world deployment of FCTS on a Boston Dynamics Spot traversing a compound parkour course consisting of a pallet, raised block, and stair descent.
Method
GPU Util.
Exp./hr
24h Exp.
Speedup
Sequential MCTS
28%
0.21
5.1
1.0×
Parallel MCTS
20%
1.13
25
5.3×
TABLE IV : Computational efficiency of parallel curriculum search.