--- tags: - robotics - lerobot - dram - rmbench --- # DRAM-PI05-RMBench This repository contains the **50,000-step RMBench checkpoint** for LeRobot π₀.₅ with DRAM streaming memory. It uses the `native_kv` mode, which takes keys and values from the vision-language model as inputs to the memory bank. ## Download ```bash hf download mystery-1057/DRAM-PI05-RMBench \ --local-dir ./checkpoints/dram-pi05-rmbench ``` Keep the checkpoint directory together. The policy loads the model, tokenizer, and observation and action processors from this directory. ## Evaluate on RMBench From the root of the DRAM-enabled LeRobot v0.6.0 checkout, with the PI dependencies installed, start the policy server: ```bash python scripts/eval_recurrent_pi05_with_rmbench.py \ --policy.path=./checkpoints/dram-pi05-rmbench \ --policy.device=cuda:0 \ --policy.dtype=float32 \ --policy.use_amp=false \ --policy.compile_model=false \ --policy.gradient_checkpointing=false \ --strict_weights=true \ --memory_tick=stride \ --observation_stride=5 \ --n_action_steps=5 \ --server.host=127.0.0.1 \ --server.port=19000 ``` Run the RMBench XPolicyLab client separately and connect to `127.0.0.1:19000`. The evaluation uses `demo_clean`, seen instructions, and 50 Hz control. The server updates DRAM memory once every five control steps and resets memory at the start of each episode.