sim-square-broad-r01-mulligan-divl

Part of Mulligan. Browse evaluations at Policy Arena; the datasets are in the mulligan organization.

State-based agent with the parent's frozen diffusion actor and a distributional DIVL critic (policy.pt, stats.json).

Property	Value
Task	sim-square-broad
Model round	R1
Arm	mulligan
Campaign cell	sq_d1_r1_ours_mining_freecf_human_only
Seeds	1, 2, 3, 4, 5 (one folder per seed)
Training step	250001

Checkpoints

Folder	Training run config
seed-1	release/run-configs/sim-square-broad-r01-mulligan-divl__seed-1.json
seed-2	release/run-configs/sim-square-broad-r01-mulligan-divl__seed-2.json
seed-3	release/run-configs/sim-square-broad-r01-mulligan-divl__seed-3.json
seed-4	release/run-configs/sim-square-broad-r01-mulligan-divl__seed-4.json
seed-5	release/run-configs/sim-square-broad-r01-mulligan-divl__seed-5.json

The frozen actor comes from sim-square-broad-r01-mulligan-idql.

Evaluations

Held-out initial-state grid; per-rollout outcomes are in the linked dataset.

Dataset	Folder	N	Successes
sim-square-broad-r00-r03-eval	seed-1	32	21271/30000
sim-square-broad-r00-r03-eval	seed-2	32	22553/30000
sim-square-broad-r00-r03-eval	seed-3	32	21085/30000
sim-square-broad-r00-r03-eval	seed-4	32	21323/30000
sim-square-broad-r00-r03-eval	seed-5	32	21384/30000

Data

Trained on: sim-square-broad-c00-teleop-sobol, sim-square-broad-c01-dagger-mulligan, sim-square-broad-c01-sobol-policy-rollouts.

Referenced by the metadata of: sim-square-broad-r00-r03-eval.

Provenance

SHA-256 of every file is recorded in release.json. .pt files are PyTorch pickles; load them only in a trusted environment. The run configs are in the Mulligan code repository; the recipes there retrain each checkpoint. License: mit.