You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

PointACT RLBench Polar Filled9 — Job 25809

This public repository contains two snapshots from the same PointACT training run.

Checkpoints

Directory	Global step	Role
`checkpoint-7500/`	7,500	Requested fixed-step snapshot
`checkpoint-8500/`	8,500	Latest complete checkpoint observed and frozen for upload on 2026-09-23

Training mode

Task: RLBench 10-task robot-action prediction
Model class: VLAEncDec3DWithActionClassificationModel
Training mode: action classification (100 position bins), not action regression
Point representation: polar filled9, 9 PointTransformer input channels
Point backbone: PointTransformer V3 / Concerto encoder mode
Frozen components: vision tower, LLM, and multimodal merger
Trainable action head position center: mixture-of-experts (moe)
Robot state input: enabled
Color augmentation: enabled
Image augmentation: enabled
Precision: BF16 with TF32 enabled; FP16 disabled
Per-device batch size: 512
Gradient accumulation steps: 1
Learning rate: 1e-4; cosine schedule; 3% warmup
Save interval: 500 steps

The “latest” label refers to the latest fully written checkpoint available when the upload snapshot was created. Training continued after the snapshot was frozen.

Downloads last month: -; Downloads are not tracked for this model. How to track

Video Preview

Robotics