Alternative
Samasource Alternatives: Managed Annotation vs Physical AI Data Capture
The best Samasource alternative depends on which half of the pipeline you are stuck on. Sama is a managed annotation vendor: you supply image, video, 3D point cloud, or text data and its workforce labels it. If your real bottleneck is capturing physical-interaction data rather than labeling data you already own, the alternative is a capture-first source. Truelabel is a physical AI data marketplace where around 10,000 collectors across 100 countries capture egocentric and teleoperation datasets with depth, IMU, and force-torque enrichment, delivered as training-ready RLDS, LeRobot, and MCAP episodes with per-trajectory provenance.
Quick facts
- Topic
- Samasource
- Audience
- Procurement leads, ML ops, robotics engineers
- Deliverable
- Buyer-facing reference + procurement guidance
What Sama does, and where the annotation model stops
Sama, formerly Samasource, is a managed data-annotation vendor. Its annotators draw bounding boxes, polygon masks, semantic segmentation, and keypoints on image, video, 3D point cloud, and text data, under project-manager review and contract SLAs [1]. The model assumes you already hold the raw data and need labeling labor at scale, and its published work centers on autonomous-vehicle perception, medical imaging, and retail recognition [2].
For a robotics or embodied-AI team, that assumption is the whole problem. The hard part of a manipulation dataset is not labeling it; it is capturing diverse real-world interaction in the first place. Sama runs no capture: no sensor rigs, no teleoperation, no multi-sensor enrichment. An alternative that only annotates leaves your scarcest input, the physical data itself, unsolved.
Sama vs Truelabel at a glance
The two platforms sit on opposite sides of the data pipeline. Sama turns data you own into labels. Truelabel is a physical AI data marketplace where you post a capture spec and vetted partners return sample datasets before you commit to scale [3]. Where that plays out in practice:
| Dimension | Sama | Truelabel |
|---|---|---|
| Core model | Managed annotation of client-supplied data | Capture-first marketplace: spec in, sampled datasets out |
| You must already own | Raw images, video, or point clouds | Nothing; collectors record net-new interaction |
| Modalities | 2D and 3D labels on images, video, LiDAR, text | Egocentric, exocentric, teleoperation, plus depth, IMU, and force-torque |
| Network | Trained annotator workforce, East Africa roots | Around 10,000 collectors across 100 countries; 100+ vetted capture partners |
| Delivery | Label files, or via Labelbox and V7 | RLDS, LeRobot, MCAP, or a custom schema with per-trajectory provenance |
| Best for | Labeling a corpus you already hold | Sourcing manipulation and teleop data you do not have yet |
Why annotation cannot retrofit embodied data
Robot policies do not learn from pixels alone. Vision-language-action models like RT-1 and RT-2 train on episodes that pair each frame with proprioceptive state and an action label [4]. Depth geometry, IMU trajectories, and force-torque readings are part of the ground truth, not annotations you can add after the fact. If a clip was recorded on a phone with no depth sensor and no gripper telemetry, no amount of bounding-box labeling recovers the contact forces a manipulation policy needs.
That is the structural limit of any annotation-only vendor. Truelabel's collectors capture the modalities at source, synchronizing RGB-D, IMU, and gripper state, and expert reviewers then label affordances, grasp types, contact events, and failure modes against that context. Failure cases carry more weight than buyers expect: grasp slips and collisions are the negative examples that make a policy robust, and guidelines written around successful detection never produce them.
Delivery format decides how fast you train
A labeled JSON manifest is not a training set for robotics. Frameworks like LeRobot and OpenVLA expect trajectory-centric episodes, with observations, actions, and rewards grouped per episode in RLDS [5]. Turning image-centric label exports into episodes means writing custom ETL, re-synchronizing sensors, and reconciling metadata by hand, which lands real work on your ML team before the first training run.
Truelabel delivers RLDS, LeRobot, and MCAP episodes with pre-aligned observation-action trajectories, calibration parameters, and metadata in the manifest, so datasets load into training scripts without a conversion layer. Each trajectory carries its consent artifact and capture metadata under Truelabel's data provenance format [6], which is what makes the data licensable and auditable rather than merely available.
Where Truelabel fits among annotation vendors
Sama competes with Appen, CloudFactory, and iMerit: same model, different logos, where you supply data and they return labels. Annotation software like Labelbox and Encord hands the tooling to your own team but still needs raw footage as input.
The closest analog to Truelabel is Scale AI's physical AI data engine, which runs custom robotics collection as project-based engagements [7]. Open corpora sit at the other end: DROID with 76,000 trajectories across 564 scenes [8], BridgeData V2 with 60,000 demonstrations [9], and Open X-Embodiment with over a million trajectories from 21 institutions [10]. They are free but fixed. You get whatever scenes the original researchers captured, with no way to commission the embodiment or task you actually deploy. A marketplace exists to fill that exact gap.
How to evaluate a Sama alternative for physical AI
Before choosing between an annotation vendor and a capture marketplace, work down from the model you are training rather than up from the cheapest per-unit price:
- 01
Separate capture from labeling
Decide whether you already own usable raw data. If you do, an annotation vendor like Sama may be enough. If the footage does not exist yet, or lacks depth, IMU, or force-torque, you need capture, and annotation cannot add those channels.
- 02
Specify embodiment and modalities
Name the robot, the sensors, and the tasks. RGB-only is cheaper; synchronized RGB-D with IMU and gripper state is what contact-rich manipulation needs. The spec, not the vendor, sets the cost.
- 03
Require a training-native format
Ask for RLDS, LeRobot, or MCAP episodes with observation-action alignment and calibration metadata. Rule out anything that forces custom ETL from image-label exports.
- 04
Check provenance and rights
Every trajectory should carry a consent artifact, a location release where applicable, and per-session metadata, so the data is licensable for commercial training and auditable later.
- 05
Buy a sample before scale
Post the spec, get a sample packet with QA evidence, and validate it against your evaluation rubric before committing to a full collection.
Ethical sourcing, two different models
Ethical sourcing sits at the center of both platforms, reached by different routes. Sama built its brand on employing annotators in East Africa with skills training and impact reporting, tracing back to its 2008 Samasource nonprofit roots. Truelabel runs a distributed collector network worldwide on task-based bounties: collectors see the payout for a defined block of data before they accept it, choose work that matches their hardware, and contribute under consent terms recorded per session. Both assign full commercial rights in the delivered data to the buyer, with no residual claim from the people who labeled or captured it. The practical difference for a buyer is upstream. Sama's workforce annotates what you send; Truelabel's collectors generate what does not exist yet.
Related pages
Use these to move from category-level context into specific task, dataset, format, and comparison detail.
External references and source context
- sama.com computer vision
Sama's computer vision annotation services cover image, video, and 3D point cloud labeling
sama.com ↩ - sama.com resources
Sama's resource library highlights case studies in autonomous vehicles, medical imaging, and retail
sama.com ↩ - truelabel physical AI data marketplace bounty intake
Truelabel operates a marketplace with around 10,000 collectors capturing physical AI datasets
truelabel.ai ↩ - RT-1: Robotics Transformer for Real-World Control at Scale
RT-1 trains on 130,000 episodes with synchronized visual observations and action labels
arXiv ↩ - RLDS: an Ecosystem to Generate, Share and Use Datasets in Reinforcement Learning
RLDS paper describes trajectory-centric dataset format for reinforcement learning
arXiv ↩ - truelabel data provenance glossary
Data provenance standards record sensor configs, timestamps, and environmental conditions
truelabel.ai ↩ - scale.com physical ai
Scale AI's physical AI data engine provides custom robotics data collection services
scale.com ↩ - Project site
DROID dataset contains 76,000 manipulation trajectories across 564 scenes and 84 tasks
droid-dataset.github.io ↩ - Project site
BridgeData V2 provides 60,000 robot manipulation demonstrations for policy learning
rail-berkeley.github.io ↩ - Open X-Embodiment: Robotic Learning Datasets and RT-X Models
Open X-Embodiment aggregates 1M+ robot trajectories from 21 institutions and 22 robot embodiments
arXiv ↩ - LeRobot documentation
LeRobot documentation describes RLDS-compatible dataset loading and training pipelines
Hugging Face - Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World
Domain randomization reduces sim-to-real gap by training on diverse visual and physical parameters
arXiv - PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation
PointNet enables deep learning on 3D point clouds for classification and segmentation
arXiv - docs.labelbox.com overview
Labelbox provides annotation platform documentation for computer vision workflows
docs.labelbox.com - v7darwin
V7 Darwin offers annotation tools and managed services for image and video datasets
v7darwin.com - MCAP specification
MCAP specification defines container format for multi-modal robotics data
MCAP - Introduction to HDF5
HDF5 introduction describes hierarchical data format for scientific datasets
The HDF Group - RLDS with TensorFlow Datasets
TensorFlow RLDS documentation explains trajectory dataset loading and preprocessing
TensorFlow - scale.com scale ai universal robots physical ai
Scale AI and Universal Robots partnership targets physical AI data for humanoid robotics
scale.com - V7 Darwin labeling services
V7 Darwin data annotation platform supports image, video, and DICOM workflows
v7darwin.com - dataloop.ai platform
Dataloop platform provides data management and annotation tools for AI workflows
dataloop.ai - MCAP guides
MCAP guides explain container format usage for robotics data recording and playback
MCAP - ROS: an open-source Robot Operating System
ROS paper describes open-source robot operating system architecture and message passing
ICRA Workshop on Open Source Software
FAQ
What types of data does Sama specialize in annotating?
Sama provides managed annotation for image, video, 3D point cloud, and text datasets, including bounding-box labeling, polygon segmentation, keypoint annotation, and semantic classification across computer vision and NLP workflows. It does not run data capture or multi-sensor enrichment, so clients must supply raw data before annotation begins.
Does Truelabel offer annotation-only services like Sama?
No. Truelabel is a capture-first physical AI data marketplace. Vetted capture partners record raw multi-modal datasets with synchronized RGB-D, IMU, and force-torque streams, and expert annotators then add semantic labels within that context. Truelabel does not take in client-provided footage for standalone annotation.
Can robotics teams use Sama for teleoperation dataset annotation?
Sama can label individual frames from teleoperation footage with bounding boxes or masks, but it does not handle trajectory-level labeling, action-sequence annotation, or multi-modal enrichment. Robotics teams need observation-action pairs, episode boundaries, and synchronized sensor streams, which sit outside an image-centric annotation model. Truelabel delivers those natively.
How long does it take to procure a custom physical AI dataset through Truelabel?
Truelabel works spec-first: you post a capture brief, receive a sample packet with QA evidence, and approve it before full collection begins. Timelines scale with the size and difficulty of the spec rather than a fixed rate card. Sama's annotation turnaround depends on task complexity and queue depth, and it assumes you already hold the raw data to label.
What delivery formats does Truelabel support for robotics training?
Truelabel delivers RLDS, LeRobot, MCAP, or a custom schema with pre-aligned observation-action trajectories. Every dataset includes sensor calibration parameters, episode metadata, and per-trajectory provenance in machine-readable manifests, so teams load datasets into LeRobot, RLDS pipelines, or custom PyTorch dataloaders without a conversion step.
Is Truelabel more expensive than Sama for equivalent data volumes?
It depends on whether you already own raw data. If you have unlabeled footage, a per-task annotation vendor may cost less than commissioning net-new capture. Most robotics teams lack diverse, sensor-rich physical data, though, and for them the real comparison is a capture marketplace versus standing up your own rigs, recruiting collectors, and building a pipeline in-house. The marketplace removes that fixed cost, and price scales with the spec you post.
Looking for samasource alternatives?
Specify modality, task, environment, requested rights posture, and delivery format. Truelabel routes the request to candidate capture partners and helps scope consent/provenance artifacts and commercial licensing requirements for buyer review before delivery.
Browse Physical AI Datasets