SplitMoE Ablation Comparison

Anchored on full (iter 27400, paper baseline). Protocol: cmd_vx = ±1.0, 20s episodes, single pit, pit_depth cap 0.72 m, stone_distance cap 0.58 m, illegal_contact disabled, 2000 envs × 13 sub-terrains × 30 rows. Excluded from filtered metrics: hurdle_pole, hurdle_board, boxes, random_rough. Full terrain cfg in baseline_terrain_config.json.
Sections: 01 success heatmap 02 routing specialization 03 velocity tracking 04 gate tsne 05 leg routing 06 wheel violin

Numerical metrics (filtered success)

Variantreach_allfwdbwdΔ vs fulltime_outmean_reward
full 0.976 0.993 0.959 — 0.024 1.08
A2 — shared critic 0.512 0.586 0.438 -0.465 0.488 1.35
A3 — no L_sym 0.812 0.776 0.846 -0.165 0.188 1.22
B1 — no L_bal 0.935 0.932 0.938 -0.041 0.065 1.09
B2 — blind 0.499 0.778 0.226 -0.478 0.501 1.62

Paper 1 — Success heatmap

full
full paper_01_success_heatmap
A2 — shared critic
A2 — shared critic paper_01_success_heatmap
A3 — no L_sym
A3 — no L_sym paper_01_success_heatmap
B1 — no L_bal
B1 — no L_bal paper_01_success_heatmap
B2 — blind
B2 — blind paper_01_success_heatmap

Paper 2 — Routing specialization

full
full paper_02_routing_specialization
A2 — shared critic
A2 — shared critic paper_02_routing_specialization
A3 — no L_sym
A3 — no L_sym paper_02_routing_specialization
B1 — no L_bal
B1 — no L_bal paper_02_routing_specialization
B2 — blind
B2 — blind paper_02_routing_specialization

Paper 3 — Velocity tracking error

full
full paper_03_velocity_tracking
A2 — shared critic
A2 — shared critic paper_03_velocity_tracking
A3 — no L_sym
A3 — no L_sym paper_03_velocity_tracking
B1 — no L_bal
B1 — no L_bal paper_03_velocity_tracking
B2 — blind
B2 — blind paper_03_velocity_tracking

Paper 4 — Combined gate t-SNE

full
full paper_04_gate_tsne
A2 — shared critic
A2 — shared critic paper_04_gate_tsne
A3 — no L_sym
A3 — no L_sym paper_04_gate_tsne
B1 — no L_bal
B1 — no L_bal paper_04_gate_tsne
B2 — blind
B2 — blind paper_04_gate_tsne

Paper 5 — Leg gate t-SNE

full
full paper_05_leg_routing
A2 — shared critic
A2 — shared critic paper_05_leg_routing
A3 — no L_sym
A3 — no L_sym paper_05_leg_routing
B1 — no L_bal
B1 — no L_bal paper_05_leg_routing
B2 — blind
B2 — blind paper_05_leg_routing

Paper 6 — Wheel expert violin

full
full paper_06_wheel_violin
A2 — shared critic
A2 — shared critic paper_06_wheel_violin
A3 — no L_sym
A3 — no L_sym paper_06_wheel_violin
B1 — no L_bal
B1 — no L_bal paper_06_wheel_violin
B2 — blind
B2 — blind paper_06_wheel_violin