JHeisler/aloha_solo_left_act_diffusion_40k

Hybrid ACT+Diffusion — ALOHA Single-Arm (Left) — 40k steps

Custom HybridACTDiffusion policy: ACT visual encoder (ResNet18 + 4-layer Transformer, mean-pooled) feeds a Diffusion U-Net decoder (FiLM conditioning, DDPM training, DDIM 10-step inference). No VAE — diffusion handles multimodal action distributions directly.

This is the 40k-step retrain (workstream S004) matching S003's step count for direct architectural comparison vs the shipped ACT-40k baseline. For the initial 13.4k baseline, see JHeisler/aloha_solo_left_act_diffusion.

Architecture

Images (cam_high, cam_left_wrist) + State (dim=9)
     │
     ▼
ACT Encoder (ResNet18 → 4-layer Transformer) → mean-pool → (B, 512) global cond vector
     │
     ▼
Diffusion U-Net (DiffusionConditionalUnet1d, FiLM modulation, down_dims=(256,512))
     │  DDPM training (100 timesteps) / DDIM 10-step inference
     ▼
Action chunks (chunk_size=100, action_dim=9)

Training Config

Field	Value
Architecture	HybridACTDiffusion (ACT encoder + Diffusion U-Net) — see lerobot/common/policies/hybrid_act_diffusion/
Dataset	JHeisler/aloha_solo_left_4_6_26 — 50 episodes, 29,785 samples, 30 fps
State / action dim	9 / 9
Cameras	cam_high, cam_left_wrist (3×480×640 each)
Steps	40,000
Batch size	28 (adaptive DOE winner — beats bs=24 by 6.8% throughput at 91.3 smpl/s)
Learning rate	3.5e-5 (linear-scaled from bs=24's 3e-5)
Total samples seen	1.12M (37 epochs over the dataset)
AMP	enabled
torch.compile	enabled
Save freq	every 10,000 steps (10k / 20k / 30k / 40k checkpoints)
Diffusion scheduler	DDPM training (100 timesteps, squaredcos_cap_v2), DDIM at inference (10 steps)
Final loss (DDPM noise-pred MSE)	0.003–0.007
Final grad norm	~0.10–0.18
Wall clock	~3h 53min on RTX A4500
LeRobot pin	96c7052777aca85d4e55dfba8f81586103ba8f61 (with custom hybrid_act_diffusion policy added)

Project Lineage

Workstream	Model	Steps	Samples	HF
S001	ACT	13,400	640K	act_left
S002	Hybrid ACT+Diffusion	13,400	321K	act_diffusion
S003	ACT (shipped)	40,000	1.92M	act_left_40k
S004	Hybrid ACT+Diffusion	40,000	1.12M	this repo

S003 vs S004 is the apples-to-apples architectural comparison: same dataset, same step count, ACT-VAE vs ACT-Diffusion decoder.

Notes on loss comparability

DDPM noise-prediction MSE (this model) and ACT's L1+KL combo (S001/S003) are different loss surfaces — absolute loss values are NOT directly comparable across architectures. The right comparison is offline action L1 on held-out episodes or real-robot rollout success rate.

Usage

# Requires lerobot pinned to 96c7052 with hybrid_act_diffusion policy package added
from lerobot.common.policies.hybrid_act_diffusion.modeling_hybrid_act_diffusion import HybridACTDiffusionPolicy
policy = HybridACTDiffusionPolicy.from_pretrained("JHeisler/aloha_solo_left_act_diffusion_40k")

Citation / Course

EN.525.681 school project — JHU Whiting School of Engineering. Team: Jake Heisler, Laura Kroening, Purushottam Shukla.

Code reference: HuggingFace LeRobot at commit 96c7052 with custom hybrid policy package.