RoboTwin 2.0-Plus

Abstract

Robustness is a prerequisite for deploying Vision-Language-Action (VLA) and World Action Model (WAM) policies beyond the conditions they were trained in. RoboTwin 2.0-Plus is the benchmark introduced in our study Do World Action Models Generalize Better than VLAs? A Robustness Study — a structured perturbation suite of seven perturbation dimensions spanning twenty-one sub-dimensions — a clean baseline plus twenty perturbation sub-dimensions — across object layout, background texture, lighting, camera viewpoint, robot initial state, language instruction, and sensor noise. Built on the RoboTwin 2.0 bimanual manipulation platform (two-arm Aloha-Agilex embodiment), every perturbation is applied per episode through a single declarative configuration and is fully backward compatible: features activate only when their keys are present, so existing configs behave identically. The result is a reproducible protocol for measuring how a policy's success degrades — one clean baseline plus seven perturbation branches over all fifty tasks.

Perturbation Suite

Seven dimensions, twenty-one sub-dimensions

Clean baseline head-camera view — **Figure 1.** Example head-camera observations for the clean baseline and each perturbation dimension. The 21-sub-dimension taxonomy is fully implemented — the clean baseline plus 20 perturbation sub-dimensions (19 active by default; C2 disabled for stability).

Sensor-noise perturbed view — **Figure 1.** Example head-camera observations for the clean baseline and each perturbation dimension. The 21-sub-dimension taxonomy is fully implemented — the clean baseline plus 20 perturbation sub-dimensions (19 active by default; C2 disabled for stability).

Usage

Quick start

Data collection uses a single command pattern. A GPU is required.

# bash collect_data.sh <task_name> <config_name> <gpu_id>

# Clean baseline — no perturbation
bash collect_data.sh beat_block_hammer demo_clean 0

# A perturbation branch, e.g. lighting (L1–L4)
bash collect_data.sh beat_block_hammer demo_light 0

Install with bash script/_install.sh, then download assets with bash script/_download_assets.sh. See the README for the full setup.

Evaluation Protocol

One baseline, seven branches

A full evaluation uses 8 configs per task: one clean baseline plus seven perturbation branches. Each branch activates exactly one dimension while holding the others at their clean defaults, isolating its effect on policy performance. Each config runs 50 episodes per task, across all 50 RoboTwin 2.0 dual-arm tasks.

Table 1. The eight evaluation configs — one clean baseline and seven single-dimension branches.
#	Config	Perturbation	Active sub-dimensions
0	`demo_clean`	None (clean baseline)	—
1	`demo_vision_noise`	Sensor Noise	N1–N5 (cycled per episode)
2	`demo_light`	Lighting	L1 (always) + L2 / L3 / L4 (stochastic)
3	`demo_camera`	Camera Viewpoints	C1 + C3 (C2 disabled by default)
4	`demo_robot_state`	Robot Initial State	Joint noise + gripper extremes
5	`demo_background_plus`	Background	B1 (texture + color tint) + B2
6	`demo_objects_plus`	Objects Layout	O1 (variable count) + O2 (pose)
7	`demo_language_plus`	Language	R1 + R2 + R3 (combined)

Perturbation taxonomy. Each dimension decomposes into sub-dimensions, individually configurable via YAML flags for combined or ablation-style evaluation. Table 2 lists all 20 perturbation sub-dimensions with implementation-level parameters.

Table 2. RoboTwin 2.0-Plus perturbation taxonomy — the 20 perturbation sub-dimensions across 7 dimensions (21 with the clean baseline), with parameter details.
Dimension	Code	Sub-dimension	Implementation details
Sensor Noise	N1	Motion blur	Gaussian kernel (r ∈ [3, 15], σ ∈ [1, 8]), rotated by a random angle ∈ [−30°, 30°]
	N2	Gaussian blur	Isotropic Gaussian (σ ∈ [1, 10])
	N3	Zoom blur	Multi-scale warp accumulation (s ∈ [1.0, 1.56], step ∈ [0.01, 0.03])
	N4	Fog	Homogeneous white fog with transmittance e^−αd (α ∈ [0.3, 1.5], d = 3)
	N5	Glass blur	Pixel displacement (δ ∈ [1, 5] px) + Gaussian (σ ∈ [0.5, 2.5]), 1–3 iterations
Light Conditions	L1	Diffuse color	Per-channel RGB tint ∈ [0.0, 3.5] applied to all directional and point lights (always on)
	L2	Direction	Spherical re-sampling: θ ∈ [8°, 82°], φ ∈ [0, 2π]; 35% chance of dramatic side lighting (θ ∈ [68°, 82°]). Linked with L4
	L3	Specular	Material specular strength ∈ [0.3, 6.0], shininess ∈ [10, 250]; all scene actors (50% chance per episode)
	L4	Shadows	Directional-light shadow on/off (50% chance); co-activated with L2
Camera Viewpoints	C1	Distance	Distance scaling ∈ [0.85, 1.0] × original (head camera only)
	C2	Spherical position	Azimuth / elevation ± 10° + ± 10% distance variation (disabled by default)
	C3	Orientation	Yaw / pitch / roll each ∈ [0°, 5°] with random sign
Robot Init States	—	Joint perturbation	Gaussian noise (std = 0.1 rad, clip ± 0.225 rad) on all joints of both arms; gripper set to extreme (0.05 or 0.95) with p = 0.25
Background Textures	B1	Scene theme	Wall + floor RGB color tint, per-channel multiplier ∈ [0.4, 1.8]; combined with upstream texture swap
Background Textures	B2	Surface appearance	Table material: metallic ∈ [0.0, 0.8], roughness ∈ [0.05, 0.95], per-channel color tint ∈ [0.4, 1.8]
Objects Layout	O1	Distractor objects	Variable count per episode (3–15 objects, vs. fixed 10 in upstream)
Objects Layout	O2	Target pose	Position: Gaussian noise (σ = 2 cm, x/y only); orientation: uniform yaw ± 15°
Language Instructions	R1	Distraction	Irrelevant conversational wrapping (~30% of combined episodes)
	R2	Common-sense rewording	Object names → functional descriptions, verb synonyms (~50%)
	R3	Reasoning chain	Imperative → goal-state / outcome description (~20%)

Seven dimensions, twenty-one sub-dimensions

Clean

Photometric noise

Light conditions

Viewpoints

Initial state

Scene textures

Layout & pose

Instructions

Quick start

One baseline, seven branches

Citing RoboTwin 2.0-Plus