Terrain Landscaping Training

No trained policy ships.

Tasks

See Mobile Robotics for the observation contract.

Targets: checked-in crater_target_manifest.json, mound_target_manifest.json; general missions are hashed and mass-balanced.

Task	Episode	Reward
Crater	300 steps (30 s)	grading MAE
Mound	300 steps (30 s)	excess volume above target
General	600 steps (60 s)	grading MAE

Documented success target (all tasks): at least 90% of observable grading cells within 0.02 m of the target for 20 consecutive accepted frames, reported at truncation. This target has not been validated against a physical system.

Choose env.external_heightmap.profile: ideal (default, no error), randomized (up to 0.1 s delay, 5% dropped frames, 2 mm elevation noise, small registration and pose errors), or failure_stale (stale-frame abort). Seeded defaults do not represent measured lab errors.

Training

Crater and mound auto-load task-local DreamerV3 and skrl PPO profiles, shadowing global profiles; general uses global profiles. See hyperparams/task/terrain_landscaping_{crater,mound}/ for these files.

srb agent train --headless --algo dreamer -e terrain_landscaping_mound env.num_envs=1
srb agent train --headless --algo skrl_ppo -e terrain_landscaping_crater env.num_envs=1

Dreamer: strict merge rejects unknown or dropped keys. Profiles use 10,000 steps and train_ratio: 512.0; gradient updates equal steps × train_ratio ÷ (batch_size × batch_length).
One environment: Dreamer landscaping rejects env.num_envs > 1; mixed batches cannot selectively reset. Stacked multi-environment particle scenes are unsupported; multi-environment particle training has not been validated.

TD-MPC2 disabled: all three tasks fail before creating logs, config, replay, or model:

TD-MPC2 landscaping is disabled: velocity-controlled shovel extension is unobserved and TD-MPC2 keeps no recurrent or history state to infer it.

GPU required: PhysX particles need a GPU; tasks read them through CPU-facing USD (sim.device = "cpu"). “Force CPU” disables particles; see Particles.

Dreamer checkpoints

Operation	Restores	Does not restore
`--continue` (same log directory)	step, agent parameters, optimizer, replay	simulator state, partial episodes, recurrent carries, RNG
Agent import into new run	agent state; starts at step 0	everything else
Evaluation or policy load	agent state for inference	everything else

Resume is not bitwise; it fails when ckpt_generation.json lacks replay chunks. Loads verify task ID, IO-schema fingerprint, observation layout, target manifest, measurement profile, and normalization against the live environment. SRB_ALLOW_UNVERIFIED_LANDSCAPING_CHECKPOINT allows missing metadata, not mismatches.

Results and limitations

Crater: no shipped controller meets the success criterion; a pushing blade disturbs the graded surface about as fast as it fills the crater.
Mound: the scripted reference clears 1.75 L.
General task: the scripted reference does not improve on zero action.
Leaving the mission-pose (±2.0 m) or body-velocity envelope terminates the episode. Out-of-envelope elevations raise ExternalHeightmapFrameError (see Modelling Envelope).
Particle behavior is simulation-only and not calibrated against regolith.

Deployment

See Sim-to-Real for bridge status, capabilities, and crater-spec validation. Validation Workflow covers dry runs.

Next: Environments