ManiLadder Models

Baseline policies for the 112 ManiLadder manipulation tasks. The repository has three top-level folders, one per method:

Folder Method What is inside
small_flow_matching_policy/ Small single-task policy: frozen RADIO backbone + transformer Flow Matching head one folder per task with the selected checkpoint and its normalization statistics (111 tasks; PressFitSoftPeg-v1 has none)
pi05/ PI0.5 finetuned with LoRA sub-folders by training scope: single task, level, suite, all tasks, plus transfer and data-scaling runs
astra/ Astra: frozen script policies written by a GPT-6 Astra coding agent (Agent-D20-E20) one folder per task with the frozen controller and its reference data (112 tasks), plus shared protocol files

Each folder has its own README with the file layout, training and evaluation details and result tables: small_flow_matching_policy/README.md, pi05/README.md, astra/README.md.

Layout

small_flow_matching_policy/<task>/      ckpt_step_<N>.pt, norm_stats.json, optional release notes
pi05/
  single_task/<task>/                   best checkpoint per task (111 tasks)
  previous_single_task/<task>/          earlier single-task runs kept for reference (62 tasks)
  level/level_<1..4>/                   one policy per difficulty level
  level_transfer/level<B>_from_level<A>/
  suite/level2_<suite>/                 one policy per level-2 suite (6 suites)
  suite_transfer/level2_<x>_from_<y>/
  suite_data_scaling/level2_rigid_n<N>/ level-2 rigid suite at 20 / 50 / 100 / 200 demos per task
  all_tasks/all112/                     one policy for all 112 tasks
astra/
  <task>/                               frozen script policy of one task
  shared/, manifest.json, results_snapshot_20261105.csv

Download part of the repository

from huggingface_hub import snapshot_download

# one PI0.5 single-task checkpoint
snapshot_download(
    repo_id="Linslab/ManiLadder_Models",
    allow_patterns="pi05/single_task/StackBlock-v1/*",
    local_dir="./models",
)

Replace the pattern with, for example, small_flow_matching_policy/StackBlock-v1/*, pi05/level/level_1/* or astra/StackBlock-v1/*. The whole repository is about 744 GB (PI0.5 431 GB, Astra 269 GB, small flow-matching 44 GB), most of it PI0.5 LoRA weights and the demonstrations inside the Astra bundles.

Repository merge (2026-11-16)

This repository now combines the former Linslab/ManiLadder_Models (small flow-matching checkpoints) and Linslab/ManiLadder_PI05_Models (PI0.5 LoRA checkpoints and frozen agent policies). The old ManiLadder_PI05_Models address redirects here. File contents were not changed; paths moved as follows:

Previous location Current path
ManiLadder_Models: <task>/... small_flow_matching_policy/<task>/...
ManiLadder_PI05_Models: single_task/... pi05/single_task/...
ManiLadder_PI05_Models: previous_single_task/... pi05/previous_single_task/...
ManiLadder_PI05_Models: multi_task/level_<N>/... pi05/level/level_<N>/...
ManiLadder_PI05_Models: multi_task/level<B>_from_level<A>/... pi05/level_transfer/level<B>_from_level<A>/...
ManiLadder_PI05_Models: multi_task/level2_<suite>/... pi05/suite/level2_<suite>/...
ManiLadder_PI05_Models: multi_task/level2_<x>_from_<y>/... pi05/suite_transfer/level2_<x>_from_<y>/...
ManiLadder_PI05_Models: multi_task/results.csv pi05/multi_task_results.csv
ManiLadder_PI05_Models: agent_d20_e20/frozen_20261102/tasks/<task>/frozen/... astra/<task>/...

Added in the merge: the PI0.5 single-task run for DualArmRubikCube-v1, the all-task policy (pi05/all_tasks/all112/), the level-2 rigid data-scaling policies (pi05/suite_data_scaling/), and twelve further Astra policies. Index files (results.csv, manifests) were updated to the new paths. Notes and per-run metadata written before the merge may still quote the previous paths, and dataset paths of the form raw/<task> or preprocessed/<task>, which are now <task>/ in the datasets linked below.

Related

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support