SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2202.10324 · NeurIPS · 2022

VRL3: A Data-Driven Framework for Visual Deep Reinforcement Learning

Che Wang, Keith Ross, Xufang Luo, Dongsheng Li

arXiv · PDF · Open in the Atlas

Code that ran

5 of 8 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; 1 honoured the contract we drafted, 0 violated it and 4 have no contract verdict; 0 failures of our instrument; 3 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/2202.10324.json · badge: /api/ran/2202.10324.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
microsoft/VRL3 canonical (classifier label)
classifier basis: classifier_2025
0 ran · 1 no recorded run
facebookresearch/drqv2 unrelated (classifier label)
classifier basis: classifier_2025
5 ran · 2 no recorded run
FunctionStatusWhere it lives
Actor Ran
ran on a synthesized input · constructed an object; no result was computed
facebookresearch/drqv2/drqv2.py
code served (permissive licence) · get_code("23bd8540206300a8")
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, Actor: 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
from torch import distributions as pyd
from torch.distributions.utils import _standard_normal

torch.manual_seed(0)

class TruncatedNormal(pyd.Normal):
    def __init__(self, loc, scale, low=-1.0, high=1.0, eps=1e-06):
        super().__init__(loc, scale, validate_args=False)
        self.low = low
        self.high = high
        self.eps = eps

    def _clamp(self, x):
        clamped_x = torch.clamp(x, self.low + self.eps, self.high - self.eps)
        x = x - x.detach() + clamped_x.detach()
        return x

    def sample(self, clip=None, sample_shape=torch.Size()):
        shape = self._extended_shape(sample_shape)
        eps = _standard_normal(shape, dtype=self.loc.dtype, device=self.loc.device)
        eps *= self.scale
        if clip is not None:
            eps = torch.clamp(eps, -clip, clip)
        x = self.loc + eps
        return self._clamp(x)

def weight_init(m):
    if isinstance(m, nn.Linear):
        nn.init.orthogonal_(m.weight.data)
        if hasattr(m.bias, 'data'):
            m.bias.data.fill_(0.0)
    elif isinstance(m, nn.Conv2d) or isinstance(m, nn.ConvTranspose2d):
        gain = nn.init.calculate_gain('relu')
        nn.init.orthogonal_(m.weight.data, gain)
        if hasattr(m.bias, 'data'):
            m.bias.data.fill_(0.0)

import types as _types
utils = _types.SimpleNamespace(TruncatedNormal=TruncatedNormal, weight_init=weight_init)

class Actor(nn.Module):
    def __init__(self, repr_dim, action_shape, feature_dim, hidden_dim):
        super().__init__()
        self.trunk = nn.Sequential(nn.Linear(repr_dim, feature_dim), nn.LayerNorm(feature_dim), nn.Tanh())
        self.policy = nn.Sequential(nn.Linear(feature_dim, hidden_dim), nn.ReLU(inplace=True), nn.Linear(hidden_dim, hidden_dim), nn.ReLU(inplace=True), nn.Linear(hidden_dim, action_shape[0]))
        self.apply(utils.weight_init)

    def forward(self, obs, std):
        h = self.trunk(obs)
        mu = self.policy(h)
        mu = torch.tanh(mu)
        std = torch.ones_like(mu) * std
        dist = utils.TruncatedNormal(mu, std)
        return dist

ARGS = {
    'repr_dim': 8,
    'action_shape': (3,),
    'feature_dim': 16,
    'hidden_dim': 32,
}
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 Actor — 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 Actor 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
Critic Ran
ran on a synthesized input · constructed an object; no result was computed
facebookresearch/drqv2/drqv2.py
code served (permissive licence) · get_code("e1f8d4dffda3351e")
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, Critic: 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
import types as _types

torch.manual_seed(0)

def weight_init(m):
    if isinstance(m, nn.Linear):
        nn.init.orthogonal_(m.weight.data)
        if hasattr(m.bias, 'data'):
            m.bias.data.fill_(0.0)
    elif isinstance(m, nn.Conv2d) or isinstance(m, nn.ConvTranspose2d):
        gain = nn.init.calculate_gain('relu')
        nn.init.orthogonal_(m.weight.data, gain)
        if hasattr(m.bias, 'data'):
            m.bias.data.fill_(0.0)

utils = _types.SimpleNamespace(**{'weight_init': weight_init})

ARGS = {
    'repr_dim': 16,
    'action_shape': (5,),
    'feature_dim': 32,
    'hidden_dim': 64
}
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 Critic — 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 Critic 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
Encoder Ran
ran on a synthesized input · constructed an object; no result was computed
facebookresearch/drqv2/drqv2.py
code served (permissive licence) · get_code("10f7e7b471d14781")
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, Encoder: 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
import types as _types

torch.manual_seed(0)

def weight_init(m):
    if isinstance(m, nn.Linear):
        nn.init.orthogonal_(m.weight.data)
        if hasattr(m.bias, 'data'):
            m.bias.data.fill_(0.0)
    elif isinstance(m, nn.Conv2d) or isinstance(m, nn.ConvTranspose2d):
        gain = nn.init.calculate_gain('relu')
        nn.init.orthogonal_(m.weight.data, gain)
        if hasattr(m.bias, 'data'):
            m.bias.data.fill_(0.0)

class Encoder(nn.Module):
    def __init__(self, obs_shape):
        super().__init__()
        assert len(obs_shape) == 3
        self.repr_dim = 32 * 35 * 35
        self.convnet = nn.Sequential(
            nn.Conv2d(obs_shape[0], 32, 3, stride=2),
            nn.ReLU(),
            nn.Conv2d(32, 32, 3, stride=1),
            nn.ReLU(),
            nn.Conv2d(32, 32, 3, stride=1),
            nn.ReLU(),
            nn.Conv2d(32, 32, 3, stride=1),
            nn.ReLU()
        )
        self.apply(utils.weight_init)
    
    def forward(self, obs):
        obs = obs / 255.0 - 0.5
        h = self.convnet(obs)
        h = h.view(h.shape[0], -1)
        return h

utils = _types.SimpleNamespace(**{n: globals()[n] for n in ['weight_init'] if n in globals()})

ARGS = {
    'obs_shape': (3, 84, 84)
}
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 Encoder — 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 Encoder 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
TruncatedNormal Ran
ran on a synthesized input · constructed an object; no result was computed
facebookresearch/drqv2/drqv2.py
code served (permissive licence) · get_code("d868560369962580")
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, TruncatedNormal: 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
from torch import distributions as pyd
from torch.distributions.utils import _standard_normal

torch.manual_seed(0)

# Create small tensors for loc and scale
loc = torch.randn(2, 4)  # batch_size=2, dim=4
scale = torch.abs(torch.randn(2, 4)) + 0.1  # ensure positive

ARGS = {
    "loc": loc,
    "scale": scale,
    "low": -1.0,
    "high": 1.0,
    "eps": 1e-06,
}
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 TruncatedNormal — 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 TruncatedNormal 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
schedule Ran
ran, honoured the contract we drafted
facebookresearch/drqv2/drqv2.py
code served (permissive licence) · get_code("e25645dfaf7e3237")
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 float, ndim 0, dtype family float, shape () 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 re

def schedule(schdl, step):
    try:
        return float(schdl)
    except ValueError:
        match = re.match('linear\\((.+),(.+),(.+)\\)', schdl)
        if match:
            init, final, duration = [float(g) for g in match.groups()]
            mix = np.clip(step / duration, 0.0, 1.0)
            return (1.0 - mix) * init + mix * final
        match = re.match('step_linear\\((.+),(.+),(.+),(.+),(.+)\\)', schdl)
        if match:
            init, final1, duration1, final2, duration2 = [float(g) for g in match.groups()]
            if step <= duration1:
                mix = np.clip(step / duration1, 0.0, 1.0)
                return (1.0 - mix) * init + mix * final1
            else:
                mix = np.clip((step - duration1) / duration2, 0.0, 1.0)
                return (1.0 - mix) * final1 + mix * final2
    raise NotImplementedError(schdl)


# Test case 1: direct float string
schdl_1 = "0.001"
step_1 = 0

# Test case 2: linear schedule
schdl_2 = "linear(0.1,0.01,100)"
step_2 = 50

# Test case 3: step_linear schedule
schdl_3 = "step_linear(0.1,0.05,50,0.01,50)"
step_3 = 75

ARGS = {
    "schdl": schdl_1,
    "step": step_1
}
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 float, ndim 0, dtype family float, shape () — shape () 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
DrQV2Agent Not yet run
unverified (harvested, no recorded run)
facebookresearch/drqv2/drqv2.py
code served (permissive licence) · get_code("20f11b6e2541fc3c")
How it ran

no invocation recorded.

eval_mode Not yet run
unverified (harvested, no recorded run)
facebookresearch/drqv2/drqv2.py
code served (permissive licence) · get_code("563be4760e7af9bc")
How it ran

no invocation recorded.

make_agent Not yet run
unverified (harvested, no recorded run)
microsoft/VRL3/src/train_adroit.py
code served (permissive licence) · get_code("f1ba293af9a573fd")
How it ran

no invocation recorded.

Also present in other papers' code

Of 8 samples lifted from this paper's linked repositories, 8 carry a function identity and 0 do not. Of the 8 distinct functions with an identity, 7 are also present in other papers whose code we hold (4 with at least one paper dated earlier, 0 the same month, 7 later) and 1 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.

FunctionAlso present inEarliest-datedOthers named
schedule
brick of this paper's assembly: helper
17 other papers
2 dated earlier, 15 later · 7 only as an assembly brick
2103.04551 (2021-03)
Behavior From the Void: Unsupervised Active Pre-Training
2107.09645 (2021-07)
Mastering Visual Continuous Control: Improved Data-Augmented Reinforcement Learning
2203.12601 (2022-03)
R3M: A Universal Visual Representation for Robot Manipulation
and 14 more not named here
eval_mode
brick of this paper's assembly: helper
9 other papers
1 dated earlier, 8 later · 9 only as an assembly brick
2103.04551 (2021-03)
Behavior From the Void: Unsupervised Active Pre-Training
2209.07364 (2022-09)
Continuous MDP Homomorphisms and Homomorphic Policy Gradient
2209.12016 (2022-09)
Mastering the Unsupervised Reinforcement Learning Benchmark from Pixels
and 6 more not named here
TruncatedNormal
brick of this paper's assembly: helper
8 other papers
1 dated earlier, 7 later · 7 only as an assembly brick
2103.04551 (2021-03)
Behavior From the Void: Unsupervised Active Pre-Training
2211.00164 (2022-11)
Principled Offline RL in the Presence of Rich Exogenous Information
2306.00975 (2023-06)
Active Vision Reinforcement Learning under Limited Visual Observability
and 5 more not named here
Encoder
brick of this paper's assembly: model
5 other papers
1 dated earlier, 4 later · 2 only as an assembly brick
2103.04551 (2021-03)
Behavior From the Void: Unsupervised Active Pre-Training
2210.10765 (2022-10)
When to Ask for Help: Proactive Interventions in Autonomous Reinforcement Learning
2310.19668 (2023-10)
DrM: Mastering Visual Reinforcement Learning through Dormant Ratio Minimization
and 2 more not named here
make_agent 4 other papers
4 later
2307.10224 (2023-07)
RL-ViGen: A Reinforcement Learning Benchmark for Visual Generalization
2312.17116 (2023-12)
Generalizable Visual Reinforcement Learning with Segment Anything Model
2402.06187 (2024-02)
Premier-TACO is a Few-Shot Policy Learner: Pretraining Multitask Representation via Temporal Action-Driven Contrastive Loss
and 1 more not named here
Critic
brick of this paper's assembly: model
2 other papers
2 later · 1 only as an assembly brick
2310.19668 (2023-10)
DrM: Mastering Visual Reinforcement Learning through Dormant Ratio Minimization
2402.12181 (2024-02)
Revisiting Data Augmentation in Deep Reinforcement Learning
Actor
brick of this paper's assembly: model
1 other paper
1 later
2310.19668 (2023-10)
DrM: Mastering Visual Reinforcement Learning through Dormant Ratio Minimization

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

We propose VRL3, a powerful data-driven framework with a simple design for solving challenging visual deep reinforcement learning (DRL) tasks. We analyze a number of major obstacles in taking a data-driven approach, and present a suite of design principles, novel findings, and critical insights about data-driven visual DRL. Our framework has three stages: in stage 1, we leverage non-RL datasets (e.g. ImageNet) to learn task-agnostic visual representations; in stage 2, we use offline RL data (e.g. a limited number of expert demonstrations) to convert the task-agnostic representations into more powerful task-specific representations; in stage 3, we fine-tune the agent with online RL. On a set of challenging hand manipulation tasks with sparse reward and realistic visual inputs, compared to the previous SOTA, VRL3 achieves an average of 780% better sample efficiency. And on the hardest task, VRL3 is 1220% more sample efficient (2440% when using a wider encoder) and solves the task with only 10% of the computation. These significant results clearly demonstrate the great potential of data-driven deep reinforcement learning.

For agents

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

get_harvested_code_for_paper("2202.10324")
get_code_for_paper("2202.10324")
have("2202.10324")

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

curl https://syntology.ai/api/ran/2202.10324.json

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

[![Syntology run record](https://syntology.ai/api/ran/2202.10324.svg)](https://syntology.ai/paper/2202.10324)

Connect an agent — have() is free.