FCTS — Failure-Guided Curriculum Tree Search Over Generated Terrain for Legged Parkour

Supplementary video

The method, the search, the results and the hardware experiments, in two minutes fifty-four.

FCTS turns each localized traversal failure into the terrain a policy trains on next, and searches over competing curricula with tree search. On held-out courses rebuilt from scanned sites it raises success from 28.4% to 85.8% on a medium course and from 14.9% to 44.7% on a hard one, and transfers to a physical quadruped.

One expansion, step by step

Validation failures drive terrain generation; parallel tree search explores alternative improvement trajectories. Step through the loop, or let it run.

The FCTS pipeline: validation traversal, failure feedback, failure-conditioned curriculum generation, rehearsal terrains, RL policy training, and the updated policy checkpoint.

Step 1 of 7

The branch the search kept

Checkpoints are nodes, edges are full training runs. PUCT balances exploitation against exploration, values back up with a maximum, and the winning branch follows the highest backed-up Q from the root. Parallel search completes 25 expansions in 24 hours, 5.3× the sequential rate.

The evaluation environments

Both environments are rebuilt from scanned geometry and held out from curriculum generation and training. Routes are fixed before evaluation and stratified by local relief, the height variation inside the robot's 0.4 m footprint.

Debris site with eleven coloured evaluation routes. — Debris site11 routes, Easy through V-Hard. Mean relief 0.14 m to 0.32 m; P95 relief reaches 0.63 m.

Exposed Beam site with two evaluation routes across steel beams. — Exposed Beam siteLarge gaps, narrow structural members and disconnected regions, where relief alone does not measure difficulty.

What the final policy does

Uncut follow-cam runs once the search has finished.

Validation terrains

The fixed set every checkpoint is scored on during the search.

Cantilever

Hanging staircase

Evaluation sites

Held-out reconstructed geometry, never used to generate curricula or train.

Debris siteFull route

Exposed ridgeNarrow supports and drops

V-Hard courseV-HardHighest relief tier

On hardware

The policy selected from the search portfolio, deployed without retuning. Control runs at 50 Hz on an onboard Jetson AGX Orin, using depth and proprioception.

Box climb and drop

Pallet, elevated block and stair descent

Scattered bricks and raised platforms

Where the baselines stop

Three regions of the Debris site. All policies start from the same checkpoint on the same route; players are synchronized, and each baseline freezes where its robot fails.

Results

Success rate on relief-stratified routes in the Debris site. Two routes per difficulty tier, each evaluated over two independent seeds with 256 trials per seed. Higher is better.
Tier	Method	Success (%)	Progress IQM	Waypoints (%)

Abstract

Quadruped parkour policies handle canonical obstacles well, but their curricula miss the irregular, compound geometry of industrial and disaster-response sites. Even training directly on reconstructed geometry from the target site does not reliably yield policies that traverse it.

FCTS uses localized traversal failures to generate the terrain a policy should train on next. Each failure's location, termination mode, robot state and surrounding geometry condition that generation, and Monte Carlo Tree Search explores alternative curriculum trajectories under a fixed budget.

On terrains built from scanned construction sites, FCTS raises success from 28.4% to 85.8% on a medium course and from 14.9% to 44.7% on a hard course, and transfers to a physical quadruped.

Cite

@article{fcts2026,
  title   = {FCTS: Failure-Guided Curriculum Tree Search Over Generated Terrain for Legged Parkour},
  author  = {TBD},
  year    = {2026}
}