ZibinDong/ActionCodec2-Extended-5mm-1k

ActionCodec2 Extended First-Order — 5mm, 1k

Trained on curated data with lower-quality datasets and most stationary-arm segments removed.

All tokenizers

Setting	Value
Motion order	1
Precision tier	5mm
Codebook budget	1024 per profile
Fitted profiles	joint, eef
Registered layouts	13
Target rate	15 Hz
Gripper	Binary, zero-order; open when ≥ 0.8

The budget applies separately to each profile; joint and EEF tokens use disjoint namespaces. Precision tiers describe physical quantization settings, not an end-to-end policy accuracy guarantee.

Usage

pip install numpy scipy torch "transformers>=4.57,<6" huggingface-hub pyyaml

import numpy as np
from transformers import AutoProcessor

codec = AutoProcessor.from_pretrained("ZibinDong/ActionCodec2-Extended-5mm-1k", trust_remote_code=True)
actions = np.zeros((12, 7), dtype=np.float32)
actions[:, 6] = 1.0

tokens = codec.encode(actions, fps=15)
decoded_actions = codec.decode(tokens, fps=15)

The default layout is single_eef_delta: position increments in metres (columns 0–2), body-frame rotation-vector increments in radians (3–5), and an absolute gripper command (6). Rotation follows R_next = R_previous @ Exp(rotvec). Supply physical values and the actual recording rate through fps; the tokenizer resamples to 15 Hz.

Inputs may have shape (T, D) or (B, T, D). Decoding returns a CPU float32 tensor of shape (B, T_out, D); resampling can change the length. Quantization is lossy.

Inspect and select layouts explicitly:

codec.print_action_spaces()
codec = codec.for_action_space("single_eef_abs")

Absolute layouts also need current_state for absolute reconstruction. Joint, EEF, single-arm and dual-arm layouts are included.

Native acceleration

The artifact bundles C++17 sources for physical quantization, Set-BPE, resampling and pose operations. Build them once to enable accelerated, parallel batch encoding and decoding:

hf download ZibinDong/ActionCodec2-Extended-5mm-1k --local-dir ./actioncodec2-tokenizer
pip install ./actioncodec2-tokenizer

A C++17 compiler is required. The processor automatically uses the installed kernels; installing the ActionCodec2 source also supplies them. Without kernels, it uses the bundled Python runtime. ACTIONCODEC2_NUM_THREADS controls the per-process thread budget; ACTIONCODEC2_NO_NATIVE=1 selects the Python reference implementation.

Keep the complete repository when downloading or saving: router_config.yaml, profiles/, runtime/, the Hugging Face entry point, metadata, native sources and license.