Architectural Designs

Please follow the following commands to reproduce the study results on vision-language policy architectures.

Reproducing Results

Fixing the ALGO to be one of [base, er, agem, ewc, packnet]. Select a policy from [bc_rnn_policy, bc_transformer_policy, bc_vilt_policy] and run:

export CUDA_VISIBLE_DEVICES=GPU_ID && \
export MUJOCO_EGL_DEVICE_ID=GPU_ID && \
python lifelong/main.py seed=SEED \
                        benchmark_name=BENCHMARK 
                        policy=POLICY \
                        lifelong=ALGO

Assitive Note - Specific command generation

Policy: Select a policy bc_rnn_policy(ResNet-RNN) bc_transformer_policy(ResNet-T) bc_vilt_policy(ViT-T)
Benchmrak: Select a benchmark libero_spatial (LIBERO-Spatial) libero_object (LIBERO-Object) libero_goal (LIBERO-Goal) libero_90 (90 short-horizon tasks in LIBERO-100) libero_10 (LIBERO-Long, 10 long-horizon tasks in LIBERO-100)
GPU Id:
Random Seed

Sequential (base)

Multitask (multitask)

Single Task (single_task)

Experience Replay (er)

Elastic Weight Consolidation (ewc)

PackNet (packnet)

Agem (agem)