You can deploy this model and start earning money today!

rangers-intrinsic/
hf_act_SC_only_connector_3cam_delta6_state16

Name: hf_act_SC_only_connector_3cam_delta6_state16
Availability: InStock

rangers-intrinsic/SC_port_72_successful_trials_simplified_delta

2304.13705

apache-2.0

Views

No views yet

Observations and Actions

This policy is an ACT model trained for SC connector insertion using 3 RGB cameras and a compact robot/task state.

Observation

At each control step, the model receives:

1. Multi-view RGB images

observation.images.left_camera
observation.images.center_camera
observation.images.right_camera

Each image has shape:

3 x 1024 x 1152

These views provide visual information about:

the cable and plug
the task board and target port
the robot end-effector relative to the insertion target

2. Low-dimensional state

observation.state has 16 dimensions:

tcp_pose.position.x
tcp_pose.position.y
tcp_pose.position.z
tcp_pose.orientation.x
tcp_pose.orientation.y
tcp_pose.orientation.z
tcp_pose.orientation.w
task.target_valid
task.cable_type_id
task.plug_type_id
task.port_type_id
task.target_module_id
task.target_port_id
task.target_module_index
task.target_port_index
task.time_limit

This state provides:

the current tool-center-point pose
numeric task conditioning describing what to insert and where

Action

The model predicts an action vector of 6 dimensions:

delta.position.x
delta.position.y
delta.position.z
delta.rotation.rx
delta.rotation.ry
delta.rotation.rz

Action semantics

This is a delta pose policy, not an absolute pose policy.

delta.position.* are Cartesian translation increments
delta.rotation.* is a rotation-vector increment

At inference time, the predicted action is applied relative to the previous commanded pose (or current pose on the first step) to produce the next Cartesian target.

In other words, the model learns to output a correction:

“move a bit in this direction”
“rotate a bit like this”

instead of directly predicting a final absolute pose.

Why delta actions?

Delta actions are often better suited for fine manipulation and insertion tasks because they:

encourage closed-loop corrective behavior
are less sensitive to absolute-frame errors
are easier to use as incremental servoing commands

This is especially useful for connector insertion, where small visual errors should produce small corrective motions rather than large jumps to an absolute pose.

Model Card for act

Action Chunking with Transformers (ACT) is an imitation-learning method that predicts short action chunks instead of single steps. It learns from teleoperated data and often achieves high success rates.

This policy has been trained and pushed to the Hub using LeRobot. See the full documentation at LeRobot Docs.

How to Get Started with the Model

For a complete walkthrough, see the training guide. Below is the short version on how to train and run inference/eval:

Train from scratch

bash

lerobot-train \
  --dataset.repo_id=${HF_USER}/<dataset> \
  --policy.type=act \
  --output_dir=outputs/train/<desired_policy_repo_id> \
  --job_name=lerobot_training \
  --policy.device=cuda \
  --policy.repo_id=${HF_USER}/<desired_policy_repo_id>
  --wandb.enable=true

Writes checkpoints to outputs/train/<desired_policy_repo_id>/checkpoints/.

Evaluate the policy/run inference