README.md

SimplerEnv (Google Robot + WidowX)

Single-arm gripper benchmarks built on SimplerEnv. Two robot families share the same simulator and venv but use different fine-tuned heads.

Variant	Robot	Tasks	Embodiment tag	HuggingFace checkpoint
SIMPLER Google	Google robot (Fractal data)	4 (Move Near, Pick Coke Can, Open/Close Drawer)	OXE_FRACTAL	RLWRLD/RLDX-1-FT-SIMPLER-GOOGLE
SIMPLER WidowX	WidowX (Bridge data)	4 (Spoon-on-Towel, Carrot-on-Plate, Stack Cube, Eggplant-in-Basket)	OXE_BRIDGE_ORIG	RLWRLD/RLDX-1-FT-SIMPLER-WIDOWX

Reported success rates: Google-VM 81.5 %, Google-VA 77.4 %, WidowX 71.9 %.

Simulator venv: rldx/eval/sim/SimplerEnv/simpler_uv/.venv.

1. Setup (one-time)

bash run_scripts/eval/simpler/setup_simpler.sh

Builds the shared SimplerEnv venv and clones ManiSkill2_real2sim.

2. Fine-tune from RLDX-1-PT

# Google (Fractal)
DATA_DIR=/path/to/fractal20220817 \
bash run_scripts/train/benchmarks/finetune_rldx1_simpler_google.sh

# WidowX (Bridge)
DATA_DIR=/path/to/bridge_orig \
bash run_scripts/train/benchmarks/finetune_rldx1_simpler_widowx.sh

Both default to RLWRLD/RLDX-1-PT as the base; override with BASE_MODEL_PATH=....

3. Run evaluation

# Google (Visual Matching)
bash run_scripts/eval/simpler/eval_simpler.sh google_vm RLWRLD/RLDX-1-FT-SIMPLER-GOOGLE

# Google (Variant Aggregation — 4 tasks × 7 variants = 28 environments)
bash run_scripts/eval/simpler/eval_simpler.sh google_va RLWRLD/RLDX-1-FT-SIMPLER-GOOGLE

# WidowX (Bridge)
bash run_scripts/eval/simpler/eval_simpler.sh widowx    RLWRLD/RLDX-1-FT-SIMPLER-WIDOWX

Outputs land in output_final/simpler_<variant>/<ckpt>/<task>/. The MODEL_PATH argument is optional and defaults to the matching released checkpoint per variant.