rsi-sorting

This repository contains the shared four-task Piper policy. The three RSI Piper repositories use the same weights; the task is selected by the observation prompt.

Required deployment configuration

Use the original training configuration through Kai0 config pi05_piper_four_tasks_sip. Model settings: pi05=True, discrete_state_input=True, action_horizon=50, internal action_dim=32, max_token_len=200.

• Use this repository's assets/piper_four_tasks/norm_stats.json for both input normalization and action denormalization.
• Effective normalization is quantile normalization. Do not substitute single-task statistics or mean/std normalization. The raw base_config defaults in training_config.json are overridden when the canonical config is resolved; effective_use_quantile_norm is true.
• Robot state and output actions have 14 dimensions: left joints 0–5, left gripper 6, right joints 7–12, right gripper 13. Model output is sliced to these 14 dimensions.
• Outputs are absolute joint/gripper targets. use_delta_joint_actions=False. Do not add the current joint state to the returned actions.
• Supply all three RGB cameras with keys top_head, hand_left and hand_right, with the same views as training.
• Preserve the original Agilex unit conventions and preprocessing. An action chunk has 50 steps; this does not require executing all 50 steps open loop.

Task prompt

Task key: sorting_dolls

The robot client must send this exact string in observation["prompt"]:

Arrange the five Russian dolls upright in a row from smallest to largest, from left to right.

The canonical list is task_prompts.json. This checkpoint does not use an "Advantage: positive" suffix.

Policy server

Download the full model repository, preserving params/ and assets/. Use the matching original source archive, pinned here:

Download matching Kai0 source code

Unzip it, enter its kai0/ directory, and install the environment using pyproject.toml and uv.lock. requirements-sip.txt records the principal tested runtime package versions.

Set CHECKPOINT to the downloaded model repository's absolute path and select an idle GPU:

export CHECKPOINT=/absolute/path/to/rsi-sorting
export CUDA_VISIBLE_DEVICES="<idle-gpu-id>"
TASK_PROMPT="$(python -c 'import json,os; print(json.load(open(os.path.join(os.environ["CHECKPOINT"],"task_prompts.json")))["sorting_dolls"])')"
python scripts/serve_policy.py \
    --port 8000 \
    --default-prompt "$TASK_PROMPT" \
    policy:checkpoint \
    --policy.config=pi05_piper_four_tasks_sip \
    --policy.dir="$CHECKPOINT"

--default-prompt only fills a missing prompt. It does not override an existing prompt from the robot client. Update any hardcoded client text to the task prompt above; changing the repository name alone does not select a task.

With the canonical loader, normalization is read from CHECKPOINT/assets/piper_four_tasks/. Do not supply a separate norm_stats override.

The package repair restores deployment assets and instructions. It does not establish a robot success rate; evaluate again after updating the actual server and robot client.