SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2407.01191 · IJCAI · 2024

MARS: Multimodal Active Robotic Sensing for Articulated Characterization

Jiahua Wang, Fang Li, Ping Zhang, Hongliang Zeng, Tingyu Ye, Chengjiong Wu

arXiv · PDF · Open in the Atlas

Code that ran

12 of 14 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (4 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 9 have no contract verdict; 3 failures of our instrument; 2 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2407.01191.json · badge: /api/ran/2407.01191.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
robhlzeng/MARS community
Papers with Code archive
12 ran · 2 no recorded run
FunctionStatusWhere it lives
FeatureFusionBlock Ran
ran on a synthesized input · constructed an object; no result was computed
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("474eb4816b8052da")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, FeatureFusionBlock: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class FeatureFusionBlock(nn.Module):
    """Transformer encoder fusion block with a learnable CLS token.

    Processes seed-point features (128 tokens) + 1 CLS token through
    a Transformer encoder for global and per-point feature extraction.
    """

    def __init__(self, d_model=512, nhead=8, num_layers=6, dim_feedforward=None, dropout=0.1):
        super().__init__()
        if dim_feedforward is None:
            dim_feedforward = 4 * d_model
        self.cls_token = nn.Parameter(torch.zeros(1, 1, d_model))
        nn.init.trunc_normal_(self.cls_token, std=0.02)
        self.pos_proj = nn.Sequential(nn.Linear(3, d_model), nn.LayerNorm(d_model))
        encoder_layer = nn.TransformerEncoderLayer(d_model=d_model, nhead=nhead, dim_feedforward=dim_feedforward, dropout=dropout, batch_first=True)
        self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
        self.norm = nn.LayerNorm(d_model)

    def forward(self, point_features, seed_xyz=None):
        """
        Args:
            point_features: (B, S, d_model) seed-point features.
            seed_xyz: optional (B, S, 3) seed-point coordinates for positional encoding.
        Returns:
            cls_feature: (B, d_model) global feature.
            point_features: (B, S, d_model) per-point features.
        """
        if seed_xyz is not None:
            point_features = point_features + self.pos_proj(seed_xyz)
        batch_size = point_features.shape[0]
        cls_tokens = self.cls_token.expand(batch_size, -1, -1)
        tokens = torch.cat([cls_tokens, point_features], dim=1)
        tokens = self.encoder(tokens)
        tokens = self.norm(tokens)
        cls_feature = tokens[:, 0, :]
        point_features = tokens[:, 1:, :]
        return (cls_feature, point_features)


ARGS = {
    "d_model": 64,
    "nhead": 4,
    "num_layers": 2,
    "dim_feedforward": 256,
    "dropout": 0.1,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type FeatureFusionBlock — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a FeatureFusionBlock output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
JointTypeDecoder Ran
ran on a synthesized input · constructed an object; no result was computed
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("483e652824478152")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, JointTypeDecoder: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# The JointTypeDecoder.__init__ takes in_dim parameter
# The forward method takes aggregated_features tensor
# aggregated_features should have shape (batch, in_dim) to pass through the MLP

# Create a small aggregated_features tensor
batch_size = 2
in_dim = 512
aggregated_features = torch.randn(batch_size, in_dim)

ARGS = {
    "in_dim": 512,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type JointTypeDecoder — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a JointTypeDecoder output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
MLDM Ran
ran on a synthesized input
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("eb00ee7e20d6b31d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.nn.functional as F

class MLDM(nn.Module):
    """Multi-Layer Dueling Module for cross-modal RGB + point cloud fusion.

    Each ResNet layer is processed into a global descriptor. Point cloud global
    features participate in the weight computation (dueling mechanism) so that
    the two modalities compete and complement each other when determining
    layer importance.

    Args:
        resnet_channels: list of channel counts for each ResNet stage.
        pc_feat_dim: point cloud per-seed-point feature dimension.
        rgb_feat_dim: intermediate RGB feature dimension per layer.
        d_model: output fused feature dimension (rgb_feat_dim + pc_feat_dim).
    """

    def __init__(self, resnet_channels, pc_feat_dim, rgb_feat_dim=256, d_model=512):
        super().__init__()
        self.num_layers = len(resnet_channels)
        self.rgb_feat_dim = rgb_feat_dim
        self.layer_processors = nn.ModuleList([nn.Sequential(nn.Conv2d(ch, rgb_feat_dim, kernel_size=1, bias=False), nn.BatchNorm2d(rgb_feat_dim), nn.ReLU(inplace=True)) for ch in resnet_channels])
        self.pc_proj = nn.Sequential(nn.Linear(pc_feat_dim, rgb_feat_dim), nn.ReLU(inplace=True))
        self.weight_mlp = nn.Sequential(nn.Linear(rgb_feat_dim * self.num_layers + rgb_feat_dim, 256), nn.ReLU(inplace=True), nn.Linear(256, self.num_layers))
        self.fusion_proj = nn.Sequential(nn.Linear(rgb_feat_dim + pc_feat_dim, d_model), nn.LayerNorm(d_model), nn.ReLU(inplace=True))

    def forward(self, rgb_features, pc_per_point):
        """
        Args:
            rgb_features: list of tensors from ResNet stages.
            pc_per_point: (B, S, C_pc) per-seed-point features from PointNet++.
        Returns:
            fused: (B, S, d_model) fused multi-modal features.
        """
        B, S, _ = pc_per_point.shape
        layer_feats = []
        for proc, feat in zip(self.layer_processors, rgb_features):
            processed = proc(feat)
            pooled = F.adaptive_avg_pool2d(processed, 1).flatten(1)
            layer_feats.append(pooled)
        layer_stack = torch.stack(layer_feats, dim=1)
        pc_global = pc_per_point.max(dim=1)[0]
        pc_global_proj = self.pc_proj(pc_global)
        rgb_concat = torch.cat(layer_feats, dim=-1)
        weight_input = torch.cat([rgb_concat, pc_global_proj], dim=-1)
        weights = F.softmax(self.weight_mlp(weight_input), dim=-1)
        rgb_global = (layer_stack * weights.unsqueeze(-1)).sum(dim=1)
        rgb_expanded = rgb_global.unsqueeze(1).expand(-1, S, -1)
        fused = torch.cat([rgb_expanded, pc_per_point], dim=-1)
        return self.fusion_proj(fused)


INIT_ARGS = {
    'resnet_channels': [8, 16],
    'pc_feat_dim': 4,
    'rgb_feat_dim': 6,
    'd_model': 10,
}

ARGS = {
    'rgb_features': [
        torch.randn(2, 8, 5, 5),
        torch.randn(2, 16, 3, 3),
    ],
    'pc_per_point': torch.randn(2, 7, 4),
}

if __name__ == '__main__':
    model = MLDM(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Expected shape: (2, 7, 10)")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 3, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
MovabilityDecoder Ran
ran on a synthesized input · constructed an object; no result was computed
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("6a155f8b4cb6ef1e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, MovabilityDecoder: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# The forward method takes point_features and passes it through an MLP
# The MLP expects input of shape (B, S, in_dim) where in_dim defaults to 512
# Output shape is (B, S, 1)

# Create a tiny point_features tensor: batch_size=2, num_points=3, feature_dim=512
point_features = torch.randn(2, 3, 512)

# The __init__ parameter is in_dim which defaults to 512
# We can use the default or specify it explicitly

ARGS = {
    "in_dim": 512,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type MovabilityDecoder — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a MovabilityDecoder output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
PerceptionScoreDecoder Ran
ran on a synthesized input
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("f4f15d43acc68ec9")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 2, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

class PerceptionScoreDecoder(nn.Module):
    """Predict perception quality score in [0, 1]."""

    def __init__(self, in_dim=512):
        super().__init__()
        self.mlp = nn.Sequential(nn.Linear(in_dim, 128), nn.LayerNorm(128), nn.ReLU(inplace=True), nn.Linear(128, 64), nn.LayerNorm(64), nn.ReLU(inplace=True), nn.Linear(64, 1), nn.Sigmoid())

    def forward(self, cls_feature):
        return self.mlp(cls_feature)


INIT_ARGS = {
    'in_dim': 4
}

ARGS = {
    'cls_feature': torch.randn(2, 4)
}

if __name__ == '__main__':
    model = PerceptionScoreDecoder(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Output range: [{output.min().item():.4f}, {output.max().item():.4f}]")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
PointNetFeaturePropagation Ran
ran on a synthesized input
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("ac744ada43a50a0e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

def index_points(points, idx):
    """Gather points by index.

    Args:
        points: (B, N, C)
        idx: (B, S) or (B, S, K)
    Returns:
        indexed: (B, S, C) or (B, S, K, C)
    """
    device = points.device
    B = points.shape[0]
    view_shape = list(idx.shape)
    view_shape[1:] = [1] * (len(view_shape) - 1)
    repeat_shape = list(idx.shape)
    repeat_shape[0] = 1
    batch_indices = torch.arange(B, dtype=torch.long, device=device).view(view_shape).repeat(repeat_shape)
    return points[batch_indices, idx, :]

def square_distance(src, dst):
    """Squared Euclidean distance between every pair of points.

    Args:
        src: (B, N, C)
        dst: (B, M, C)
    Returns:
        dist: (B, N, M)
    """
    return torch.cdist(src, dst, p=2.0).pow(2)

class PointNetFeaturePropagation(nn.Module):
    """Feature Propagation module: inverse-distance interpolation + MLP."""

    def __init__(self, in_channel, mlp_channels):
        super().__init__()
        layers = []
        last_channel = in_channel
        for out_channel in mlp_channels:
            layers.append(nn.Conv1d(last_channel, out_channel, 1, bias=False))
            layers.append(nn.BatchNorm1d(out_channel))
            layers.append(nn.ReLU(inplace=True))
            last_channel = out_channel
        self.mlp = nn.Sequential(*layers)

    def forward(self, xyz1, xyz2, features1, features2):
        """Propagate features from xyz2 (fewer points) to xyz1 (more points).

        Args:
            xyz1: (B, N, 3) target points (more points)
            xyz2: (B, S, 3) source points (fewer points)
            features1: (B, N, C1) skip-connection features at target level
            features2: (B, S, C2) features to propagate from source level
        Returns:
            new_features: (B, N, C_out)
        """
        B, N, _ = xyz1.shape
        _, S, _ = xyz2.shape
        if S == 1:
            interpolated = features2.expand(-1, N, -1)
        else:
            dists = square_distance(xyz1, xyz2)
            dists, idx = dists.topk(3, dim=-1, largest=False)
            dist_recip = 1.0 / (dists + 1e-08)
            weight = dist_recip / dist_recip.sum(dim=-1, keepdim=True)
            interpolated = torch.sum(index_points(features2, idx) * weight.unsqueeze(-1), dim=2)
        if features1 is not None:
            new_features = torch.cat([features1, interpolated], dim=-1)
        else:
            new_features = interpolated
        new_features = new_features.permute(0, 2, 1)
        new_features = self.mlp(new_features)
        new_features = new_features.permute(0, 2, 1)
        return new_features

INIT_ARGS = {
    'in_channel': 8,
    'mlp_channels': [16, 12]
}

ARGS = {
    'xyz1': torch.randn(2, 12, 3),
    'xyz2': torch.randn(2, 5, 3),
    'features1': torch.randn(2, 12, 4),
    'features2': torch.randn(2, 5, 4)
}

if __name__ == '__main__':
    model = PointNetFeaturePropagation(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Expected shape: (2, 12, 12)")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 3, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
PointNetSetAbstraction Ran
ran on a synthesized input · constructed an object; no result was computed
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("bf8d6ab94618b20c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, PointNetSetAbstraction: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    "npoint": 2,
    "nsample": 3,
    "in_channel": 5,
    "mlp_channels": [8, 16]
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type PointNetSetAbstraction — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a PointNetSetAbstraction output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
ResNetBackbone Ran
ran on a synthesized input
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("d452cf2b41931a14")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned list. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
from torchvision.models import resnet18

class ResNetBackbone(nn.Module):
    """ResNet18 backbone that extracts multi-layer features.

    Returns features from 5 stages of ResNet18:
      - layer0 (conv1+bn1+relu+maxpool): (B, 64, H/4, W/4)
      - layer1: (B, 64, H/4, W/4)
      - layer2: (B, 128, H/8, W/8)
      - layer3: (B, 256, H/16, W/16)
      - layer4: (B, 512, H/32, W/32)
    """

    def __init__(self, pretrained=True):
        super().__init__()
        if pretrained:
            try:
                from torchvision.models import ResNet18_Weights
                backbone = resnet18(weights=ResNet18_Weights.DEFAULT)
            except:
                backbone = resnet18(weights=None)
        else:
            backbone = resnet18(weights=None)
        self.layer0 = nn.Sequential(backbone.conv1, backbone.bn1, backbone.relu, backbone.maxpool)
        self.layer1 = backbone.layer1
        self.layer2 = backbone.layer2
        self.layer3 = backbone.layer3
        self.layer4 = backbone.layer4

    def forward(self, x):
        """
        Args:
            x: (B, 3, H, W) RGB image tensor.
        Returns:
            List of 5 feature tensors from each stage.
        """
        f0 = self.layer0(x)
        f1 = self.layer1(f0)
        f2 = self.layer2(f1)
        f3 = self.layer3(f2)
        f4 = self.layer4(f3)
        return [f0, f1, f2, f3, f4]


INIT_ARGS = {
    'pretrained': False
}

ARGS = {
    'x': torch.randn(2, 3, 64, 64)
}

if __name__ == '__main__':
    model = ResNetBackbone(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Number of feature maps: {len(output)}")
    for i, feat in enumerate(output):
        print(f"Feature {i} shape: {feat.shape}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch, torchvision — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list — a list output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
VotingModule Ran
ran on a synthesized input
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("12732a2ba582423d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.nn.functional as F

class VotingModule(nn.Module):
    """Per-seed-point voting for joint parameters.

    Each seed point independently votes for the joint axis position,
    direction, and state. Final predictions are the mean of all votes.
    No hard thresholding is used, avoiding gradient issues during training.
    """

    def __init__(self, in_dim=512, hidden_dim=256):
        super().__init__()
        self.feat_offset = nn.Sequential(nn.Linear(in_dim, hidden_dim), nn.ReLU(inplace=True), nn.Linear(hidden_dim, in_dim))
        self.vote_h = nn.Sequential(nn.Linear(in_dim, hidden_dim), nn.ReLU(inplace=True), nn.Linear(hidden_dim, 3))
        self.vote_u = nn.Sequential(nn.Linear(in_dim, hidden_dim), nn.ReLU(inplace=True), nn.Linear(hidden_dim, 3))
        self.vote_v = nn.Sequential(nn.Linear(in_dim, hidden_dim), nn.ReLU(inplace=True), nn.Linear(hidden_dim, 1))

    def forward(self, seed_xyz, seed_features):
        """
        Args:
            seed_xyz: (B, S, 3) seed point coordinates.
            seed_features: (B, S, C) seed point features.
        Returns:
            h: (B, 3) joint axis position (mean of per-point votes).
            u: (B, 3) joint axis direction (normalized mean vote).
            v: (B, 1) joint state.
            agg_features: (B, C) aggregated features for downstream.
        """
        offsets = self.feat_offset(seed_features)
        voted_features = seed_features + offsets
        h_offsets = self.vote_h(voted_features)
        h_votes = seed_xyz + h_offsets
        u_votes = self.vote_u(voted_features)
        v_votes = self.vote_v(voted_features)
        h = h_votes.mean(dim=1)
        u = F.normalize(u_votes.mean(dim=1), dim=-1)
        v = v_votes.mean(dim=1)
        agg_features = voted_features.mean(dim=1)
        return (h, u, v, agg_features)


INIT_ARGS = {
    'in_dim': 4,
    'hidden_dim': 8,
}

ARGS = {
    'seed_xyz': torch.randn(2, 5, 3),
    'seed_features': torch.randn(2, 5, 4),
}

if __name__ == '__main__':
    module = VotingModule(**INIT_ARGS)
    h, u, v, agg_features = module(**ARGS)
    print(f"h shape: {h.shape}, expected: (2, 3)")
    print(f"u shape: {u.shape}, expected: (2, 3)")
    print(f"v shape: {v.shape}, expected: (2, 1)")
    print(f"agg_features shape: {agg_features.shape}, expected: (2, 4)")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
farthest_point_sample Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("9735279d0bc3aa74")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family int, shape (2, 2) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create a small batch of point clouds
# xyz shape: (B, N, 3) where B=batch size, N=number of points, 3=coordinates
B = 2
N = 4
xyz = torch.randn(B, N, 3, dtype=torch.float32)

# Number of points to sample
npoint = 2

ARGS = {
    "xyz": xyz,
    "npoint": npoint,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family int, shape (2, 2) — shape (2, 2) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
index_points Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("947dcee3f577f6cd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 4, dtype family float, shape (2, 4, 2, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create small tensors for the function
B = 2  # batch size
N = 5  # number of points
C = 3  # number of channels
S = 4  # number of sampled points
K = 2  # number of neighbors

# points: (B, N, C) - batch of point clouds
points = torch.randn(B, N, C)

# idx: (B, S, K) - indices to gather
idx = torch.randint(0, N, (B, S, K))

ARGS = {
    "points": points,
    "idx": idx,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 4, 2, 3) — shape (2, 4, 2, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
square_distance Ran
ran; our contract draft was wrong, not the code
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("24896468cac00905")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 3, dtype family float, shape (2, 3, 5) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create small tensors with distinct dimensions
# src: (B=2, N=3, C=4)
# dst: (B=2, M=5, C=4)
src = torch.randn(2, 3, 4)
dst = torch.randn(2, 5, 4)

ARGS = {
    "src": src,
    "dst": dst,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 3, dtype family float, shape (2, 3, 5) — shape (2, 3, 5) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
MFFP Not yet run
unverified (harvested, no recorded run)
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("0cda0dd251868fc7")
How it ran

no invocation recorded.

PointNet2Backbone Not yet run
unverified (harvested, no recorded run)
robhlzeng/MARS/mars/models/mffp.py
code served (permissive licence) · get_code("5975f43c484850be")
How it ran

no invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 14 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Precise perception of articulated objects is vital for empowering service robots. Recent studies mainly focus on point cloud, a single-modal approach, often neglecting vital texture and lighting details and assuming ideal conditions like optimal viewpoints, unrepresentative of real-world scenarios. To address these limitations, we introduce MARS, a novel framework for articulated object characterization. It features a multi-modal fusion module utilizing multi-scale RGB features to enhance point cloud features, coupled with reinforcement learning-based active sensing for autonomous optimization of observation viewpoints. In experiments conducted with various articulated object instances from the PartNet-Mobility dataset, our method outperformed current state-of-the-art methods in joint parameter estimation accuracy. Additionally, through active sensing, MARS further reduces errors, demonstrating enhanced efficiency in handling suboptimal viewpoints. Furthermore, our method effectively generalizes to real-world articulated objects, enhancing robot interactions. Code is available at https://github.com/robhlzeng/MARS.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2407.01191")
get_code_for_paper("2407.01191")
have("2407.01191")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2407.01191.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2407.01191.svg)](https://syntology.ai/paper/2407.01191)

Connect an agent — have() is free.