abot-physworld-physics-aligned-world-model

name	abot-physworld-physics-aligned-world-model
title	ABot-PhysWorld: 14B DiT for Physics-Aligned Robotic Manipulation Videos
version	0.0.3
engine	skillxiv-v0.0.3-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2603.23376
keywords	["Physics Alignment","Diffusion Transformer","Robotic Manipulation","DPO Training","Video Generation"]
description	Replace standard likelihood-based video diffusion training with decoupled physics discriminators and DPO post-training to suppress physically implausible behaviors (object penetration, anti-gravity motion) in robotic manipulation videos. Use when generating physics-realistic video predictions for embodied AI and want to maintain visual quality without physical violations.
category	Component Innovation

What This Skill Does

Replace standard diffusion model training objectives with a decoupled discriminator architecture and DPO (Direct Preference Optimization) post-training. This eliminates physically implausible behaviors like object interpenetration and anti-gravity motion while maintaining visual quality in robotic manipulation video generation.

The Component Swap

The old approach uses a single likelihood-based diffusion loss that prioritizes pixel-level reconstruction without explicit physics constraints:

# Old: standard diffusion training
# Loss = prediction_error(video_pred, video_real)
# No explicit physics modeling

The new ABot-PhysWorld approach decouples physics supervision from visual supervision using separate discriminator heads:

# New: decoupled discriminators for physics and visual quality
# Physics discriminator: classifies trajectory plausibility
physics_discriminator = DiscriminatorHead(
    input_dim=latent_dim,
    output_dim=1,  # Binary: physically plausible or not
    detects=['interpenetration', 'anti_gravity', 'contact_violations']
)

# Visual discriminator: classifies visual quality
visual_discriminator = DiscriminatorHead(
    input_dim=latent_dim,
    output_dim=1  # Binary: visually realistic or not
)

# DPO loss combines both signals
loss_dpo = dpo_loss(
    preferred=model_output_physics_plausible,
    rejected=model_output_physics_implausible,
    beta=0.1  # Preference strength
)

Post-training with DPO uses curated preference pairs to reinforce physically plausible trajectories while downweighting implausible ones. Parallel context blocks inject spatial action information without modifying core diffusion parameters:

# Parallel context injection for robotic action conditioning
action_embedding = embed(action_tokens)  # Action history
context_block = ParallelContextAttention(
    context=action_embedding,
    video_features=diffusion_features
)
# Applied in parallel to main diffusion UNet, not sequentially

Performance Impact