File size: 1,400 Bytes
8892eba
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
---
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.