Farpoint-ACT-V0

Model release: v0.0.0
Training checkpoint: step-200000
Training dataset: wenyixu101/so101-sim-oracle-pick-and-place, revision v0.1.4
Status: Experimental, simulation-only

Farpoint-ACT-V0 is an ACT policy trained for a simulated SO-101 cube pick-and-place task using synchronized front and wrist camera observations. This first release is published as a reproducible project baseline, not as a production or real-robot model.

Released checkpoint

The repository root contains the LeRobot-compatible checkpoint trained for 200,000 optimizer steps.

Field	Value
Model release	v0.0.0
Checkpoint	step-200000
Policy	ACT
Model SHA256	fc444f76dd61bd4cf9b982c4e93ff406800e713be59542499e6e18fe85474a83
Training code commit	c6e59c035d3cdae9470acaf8d955e6069ff7ea2c
Training image	sha256:d99274c14bc7e1064f3ad534deb1feecdcbeb271c9d04b3e46377d464c720293

Training data

The model was trained on wenyixu101/so101-sim-oracle-pick-and-place@v0.1.4:

• Resolved dataset commit: c137559a9e5fa39bf9db82d0c35e345cf75ed060
• Training episodes: 0:260
• Validation episodes: 260:280
• 200 nominal demonstrations
• 40 approach-stage recovery demonstrations
• 20 grasp-stage recovery demonstrations
• 20 transport-stage recovery demonstrations
• Legacy pre-lift recovery demonstrations excluded

Balanced recovery sampling was used throughout training:

• Nominal: 80%
• Approach recovery: 6.67%
• Grasp recovery: 6.67%
• Transport recovery: 6.67%
• Blue and red objects were sampled equally within every group

Autonomous rollout evaluation

The released checkpoint was evaluated on a frozen, independent 20-scene simulation holdout. Collection and recovery scenes were excluded. All evaluated episodes contain synchronized front and wrist videos.

Evaluation control:

• Maximum 1,200 policy steps
• Replanning interval: 10 steps
• Front and wrist camera observations
• so101-viam-50deg-s-v1 action-safety profile

Metric	Result
Task success	9/20 (45%)
Blue success	5/10
Red success	4/10
Cube contact	20/20
Bilateral contact	18/20
Stable grasp	17/20
Lift	16/20
Target entry	10/20
Released after lift	11/20
Stable release	9/20
Non-finite actions	0

Terminal outcomes were 9 successes, 6 lifts without target entry, 4 contacts without lift, and 1 target entry without stable release.

Action-safety observations:

• Delta-limited commands: 2,404
• Raw hard-range violations clipped before execution: 751
• Maximum raw hard-range excess: 4.603 calibrated units
• Non-finite actions: 0

Training-length ablation

The original 200k checkpoint was continued to 300k using the same dataset, image, sampler, seed, and control configuration. Lower teacher-forced validation loss did not translate into higher closed-loop success.

Checkpoint	Selection note	Frozen-20 success	Blue	Red
200k	Released checkpoint	9/20	5/10	4/10
240k	Best new teacher-forced validation loss	8/20	6/10	2/10
300k	Final continuation checkpoint	5/20	4/10	1/10

The 200k checkpoint is released because it performed best under autonomous closed-loop rollout, despite the 240k checkpoint having the lowest teacher-forced validation loss among the continuation checkpoints.

Intended use

This model is intended for:

• Reproducing the Farpoint simulated SO-101 ACT baseline
• Studying recovery-balanced imitation-learning datasets
• Comparing training checkpoints under frozen autonomous rollouts
• Research and educational experiments in simulation

Limitations

• The model was trained and evaluated only in simulation.
• This release does not demonstrate sim-to-real transfer or real-robot safety.
• The dataset and 20-scene evaluation holdout are small.
• Performance is uneven across object colors and physical variants.
• Overall autonomous success is 45%.
• The action limiter materially affects closed-loop behavior.
• Lower teacher-forced validation loss was not a reliable predictor of rollout success.
• This checkpoint is not suitable for unattended real-robot deployment.

Loading with LeRobot

from lerobot.policies.act.modeling_act import ACTPolicy

policy = ACTPolicy.from_pretrained(
    "wenyixu101/Farpoint-ACT-V0",
)

Use the same observation keys, normalization processors, action ordering, camera setup, and control limits as the training environment.