SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2307.08243 · ICCV · 2023

Uncertainty-aware State Space Transformer for Egocentric 3D Hand Trajectory Forecasting

Junsong Yuan, Yi Xu, Zhong Li, Yu Kong, Wentao Bao, Lele Chen, Libing Zeng

arXiv · PDF · Open in the Atlas

Code that ran

9 of 13 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 9 failures of our instrument; 4 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/2307.08243.json · badge: /api/ran/2307.08243.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
oppo-us-research/USST canonical (classifier label)
classifier basis: census
9 ran · 4 no recorded run
FunctionStatusWhere it lives
denormalize_traj Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oppo-us-research/USST/src/H2OLoader.py
code served (permissive licence) · get_code("dad48a20c072ff6e")
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 ndarray, ndim 2, dtype family float, shape (4, 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 numpy as np

# Create normalized trajectory data
# For '3d' target: 3D coordinates (x, y, z)
traj_norm_3d = np.random.rand(4, 3).astype(np.float32)

# For '2d' target: 2D coordinates (x, y)
traj_norm_2d = np.random.rand(5, 2).astype(np.float32)

ARGS = {
    'traj_norm': traj_norm_3d,
    'target': '3d',
    'use_global': True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family float, shape (4, 3) — shape (4, 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
draw_uv_points Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oppo-us-research/USST/src/utils_vis.py
code served (permissive licence) · get_code("681c38dc9a4a43ce")
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 ndarray, ndim 4, dtype family int, shape (2, 480, 640, 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 numpy as np
import cv2

# Create a small video tensor: (T, H, W, C)
# T=2 frames, H=480, W=640, C=3 (RGB)
video = np.zeros((2, 480, 640, 3), dtype=np.uint8)

# Create predictions: (Tm, 2/3) where Tm is number of predictions
# Using 2 predictions with 2D coordinates (u, v normalized to [0,1])
preds = np.array([
    [0.5, 0.5],      # center of frame
    [0.25, 0.75]     # quarter position
], dtype=np.float32)

# Define ARGS dictionary with all parameters
ARGS = {
    "video": video,
    "preds": preds,
    "start": 0,
    "color": (0, 255, 0)
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
cv2, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 4, dtype family int, shape (2, 480, 640, 3) — shape (2, 480, 640, 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
get_depth_anchors Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oppo-us-research/USST/src/utils.py
code served (permissive licence) · get_code("a0bee2e8ddb51a58")
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 float, shape (1, 1) 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 numpy as np
import torch

torch.manual_seed(0)

# Create trajectory ground truth as a tensor
# Shape: (num_timesteps, 3) where last dimension has x, y, z coordinates
traj_gt = torch.tensor([
    [1.0, 2.0, 5.0],
    [1.5, 2.5, 6.0],
    [2.0, 3.0, 7.0],
    [2.5, 3.5, 8.0],
], dtype=torch.float32)

num_obs = 2  # number of observed timesteps
num_full = 4  # total number of timesteps

# Create time_orders matrix for polynomial fitting
# Shape: (num_full, polynomial_degree)
# This is used in lstsq for least squares fitting
time_orders = torch.tensor([
    [1.0, 0.0],
    [1.0, 1.0],
    [1.0, 2.0],
    [1.0, 3.0],
], dtype=torch.float32)

offset_anchor = 'Zc'

ARGS = {
    'traj_gt': traj_gt,
    'num_obs': num_obs,
    'num_full': num_full,
    'offset_anchor': offset_anchor,
    'time_orders': time_orders,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, 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, shape (1, 1) — shape (1, 1) 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
get_optimizer Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oppo-us-research/USST/src/optimizers.py
code served (permissive licence) · get_code("1df9d391050a4608")
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 SGD. 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)

# Create a minimal config object with required attributes
class Config:
    def __init__(self, optimizer='SGD', base_lr=0.01, momentum=0.9, weight_decay=1e-4, betas=(0.9, 0.999)):
        self.optimizer = optimizer
        self.base_lr = base_lr
        self.momentum = momentum
        self.weight_decay = weight_decay
        self.betas = betas

# Create a minimal neural network module
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)
    
    def forward(self, x):
        return self.linear(x)

# Instantiate the model and get its parameters
model = SimpleModel()
params = model.parameters()

# Create config for SGD optimizer
cfg = Config(optimizer='SGD', base_lr=0.01, momentum=0.9)

# Define arguments dictionary
ARGS = {
    'cfg': cfg,
    'params': params
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
adan_pytorch, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type SGD — a SGD output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a SGD 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
get_scheduler Ran
ran; our contract draft was wrong, not the code
oppo-us-research/USST/src/optimizers.py
code served (permissive licence) · get_code("6652cb820c60c46b")
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 CosineAnnealingWarmupRestarts_New. 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 math

torch.manual_seed(0)

# Create a minimal config object with all required attributes
class Config:
    def __init__(self):
        self.scheduler = SchedulerConfig()
        self.epoch = 100
        self.base_lr = 0.1

class SchedulerConfig:
    def __init__(self):
        self.type = 'cosine_warmup'
        self.warmup_epoch = 5
        self.lr_decay = 0.9
        self.lr_decay_epoch = [50, 75]
        self.num_restarts = 1
        self.min_lr_factor = 0.01

# Create a simple model with parameters to optimize
class SimpleModel(nn.Module):
    def __init__(self):
        super(SimpleModel, self).__init__()
        self.linear = nn.Linear(4, 3)
    
    def forward(self, x):
        return self.linear(x)

# Create model and optimizer
model = SimpleModel()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)

# Create config
cfg = Config()

# Define ARGS
ARGS = {
    'cfg': cfg,
    'optimizer': optimizer
}
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 CosineAnnealingWarmupRestarts_New — a CosineAnnealingWarmupRestarts_New output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a CosineAnnealingWarmupRestarts_New 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
load_checkpoint Ran
ran; our contract draft was wrong, not the code
oppo-us-research/USST/src/utils_io.py
code served (permissive licence) · get_code("6039e11b8c330d4d")
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 tuple, ndim 1, 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 os
import tempfile
import torch
import torch.nn as nn
import numpy as np

torch.manual_seed(0)

# Create a temporary directory structure for checkpoint files
temp_dir = tempfile.mkdtemp()
dataset_name = "test_dataset"
tag_name = "test_tag"
ckpt_dir = os.path.join(temp_dir, dataset_name, tag_name, 'snapshot')
os.makedirs(ckpt_dir, exist_ok=True)

# Create a minimal model
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)
    
    def forward(self, x):
        return self.linear(x)

# Create and save checkpoint files
model = MinimalModel()
checkpoint_data = {'model': model.state_dict()}

# Save two checkpoint files to test the filtering logic
checkpoint_path_01 = os.path.join(ckpt_dir, 'snapshot_epoch_01.pth')
checkpoint_path_02 = os.path.join(ckpt_dir, 'snapshot_epoch_02.pth')
torch.save(checkpoint_data, checkpoint_path_01)
torch.save(checkpoint_data, checkpoint_path_02)

# Create a config object with all required attributes
class Config:
    def __init__(self):
        self.output_dir = temp_dir
        self.tag = tag_name
        self.DATA = type('obj', (object,), {'dataset': dataset_name})()
        self.TEST = type('obj', (object,), {'test_epoch': None})()
        self.TRAIN = type('obj', (object,), {'snapshot_prefix': 'snapshot_epoch_'})()

cfg = Config()

# Create a fresh model instance for loading
model_for_loading = MinimalModel()

ARGS = {
    'cfg': cfg,
    'model': model_for_loading
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
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
normalize_traj Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oppo-us-research/USST/src/H2OLoader.py
code served (permissive licence) · get_code("8da2effca700d6c7")
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 ndarray, ndim 2, dtype family float, shape (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 numpy as np

# Create test trajectory data
# For '3d' target: need shape (N, 3) with values in the range of H2O_STATS
traj_3d = np.array([
    [0.0, 0.0, 0.4],
    [0.2, 0.1, 0.5],
], dtype=np.float32)

# For '2d' target: need shape (N, 2) with values in pixel range
traj_2d = np.array([
    [640.0, 360.0],
    [800.0, 400.0],
], dtype=np.float32)

ARGS = {
    'traj': traj_3d,
    'target': '3d',
    'use_global': True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family float, shape (2, 3) — shape (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
send_to_gpu Ran
ran; our contract draft was wrong, not the code
oppo-us-research/USST/src/utils.py
code served (permissive licence) · get_code("4b07133043aed642")
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 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

torch.manual_seed(0)

# Create a tuple of small tensors
tensor1 = torch.randn(2, 4)
tensor2 = torch.randn(2, 3)
tensor3 = torch.randn(2, 5)

data_tuple = (tensor1, tensor2, tensor3)

# Device parameter - use CPU only
device = torch.device('cpu')

# non_blocking parameter
non_blocking = False

ARGS = {
    'data_tuple': data_tuple,
    'device': device,
    'non_blocking': non_blocking
}
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 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: none by the harness; any seed is in the setup block
world_to_camera Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oppo-us-research/USST/src/H2OLoader.py
code served (permissive licence) · get_code("9b2d18e6c9ff858a")
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 ndarray, ndim 2, dtype family float, shape (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 numpy as np

# Analyze the function:
# - traj3d_global: iterable of xyz points (each xyz is shape (3,) based on indexing)
# - cam2world: iterable of Rc2w matrices (each is 4x4 based on inversion and matrix multiplication)
# The function zips over these two iterables

# Create sample 3D trajectory points in global coordinates
# xyz should be shape (3,) for each point
xyz_point_1 = np.array([1.0, 2.0, 3.0])
xyz_point_2 = np.array([4.0, 5.0, 6.0])
traj3d_global = [xyz_point_1, xyz_point_2]

# Create sample camera-to-world transformation matrices
# Rc2w should be 4x4 (homogeneous transformation matrices)
# Identity matrices as simple stand-ins
cam2world_1 = np.eye(4)
cam2world_2 = np.array([
    [0.9, -0.1, 0.0, 0.5],
    [0.1, 0.9, 0.0, 1.0],
    [0.0, 0.0, 1.0, 2.0],
    [0.0, 0.0, 0.0, 1.0]
])
cam2world = [cam2world_1, cam2world_2]

ARGS = {
    "traj3d_global": traj3d_global,
    "cam2world": cam2world
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family float, shape (2, 3) — shape (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
build_dataloaders Not yet run
unverified (harvested, no recorded run)
oppo-us-research/USST/src/EgoPAT3DLoader.py
code served (permissive licence) · get_code("92f4580f0daea9e6")
How it ran

no invocation recorded.

gather_eval_results Not yet run
unverified (harvested, no recorded run)
oppo-us-research/USST/src/utils.py
code served (permissive licence) · get_code("aa4989cb736e5e1f")
How it ran

no invocation recorded.

parse_configs Not yet run
unverified (harvested, no recorded run)
oppo-us-research/USST/src/config.py
code served (permissive licence) · get_code("169badeb8cb30b4a")
How it ran

no invocation recorded.

vis_demo Not yet run
unverified (harvested, no recorded run)
oppo-us-research/USST/src/utils_vis.py
code served (permissive licence) · get_code("b09593fc01e9188e")
How it ran

no invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 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 13 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

Egocentric 3D Hand Trajectory Forecasting. Our goal is to predict the future 3D hand trajectory (in red) given the past observations of egocentric video and trajectory (in blue). Compared to the 2D image space, predicting trajectory in a global 3D space is practically more valuable to understand human intention and behavior in AR/VR applications.

For agents

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

get_harvested_code_for_paper("2307.08243")
get_code_for_paper("2307.08243")
have("2307.08243")

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

curl https://syntology.ai/api/ran/2307.08243.json

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

[![Syntology run record](https://syntology.ai/api/ran/2307.08243.svg)](https://syntology.ai/paper/2307.08243)

Connect an agent — have() is free.