FAM

Few-Shot Spatial Intelligence for Embodied Manipulation

FAM preserves task-relevant spatial relationships across vision, language and action to support rapid adaptation under defined conditions.

Embodied manipulation foundation model

FAM: A Few-Shot, End-to-End Foundation Model for Embodied Manipulation

Adapt to new tasks, environments and robot platforms with only a small number of real-world demonstrations.

Before

Traditional VLA Representation

  • 3D visual information compressed into a one-dimensional latent vector
  • Spatial relationships lost during representation compression
  • Limited interpretability
  • Information bottleneck

After

FAM Spatial Representation

  • Two-dimensional and three-dimensional heatmap representation
  • Spatial alignment across vision, language and action
  • Preserved object relationships
  • Interpretable navigation and manipulation cues
01

Spatial Understanding Pretraining

Learning spatial keypoints, object relationships, interaction patterns and motion trajectories from large-scale human operation videos.

02

3D Few-Shot Alignment

Aligning visual-language representations and action policies at the spatial heatmap level.

03

Rapid Real-Robot Adaptation

Adapting typical tasks with only a small number of real-world demonstrations, subject to task complexity, hardware configuration and environmental conditions.

RLBenchPeg InsertionDrawer OpeningLighting VariationBackground InterferenceObstacle Interference

Measured under defined conditions

Performance Designed to Be Verifiable

Every product specification and performance claim is linked to configurable test conditions, evidence, evaluation methods and limitations.

88.2%

RLBench Simulation Manipulation Success Rate

97%

Defined Base-Task Success Rate in Complex Real Environments

3–5

Real Demonstrations for Typical Task Adaptation

1%

Data Requirement Compared with Selected Traditional Pipelines

28 DoF

Full-Body Degrees of Freedom

24/7

Continuous Operation with Dual Hot-Swap Batteries

Validation details

88.2% · RLBench Simulation Manipulation Success Rate

Test Conditions
Selected manipulation tasks under a controlled simulation setup.
Dataset
RLBench task suite; exact task subset is configurable in the validation record.
Task Definition
Complete the specified manipulation goal within the evaluation horizon.
Evaluation Method
Task-level success averaged across the configured benchmark subset.
Robot Configuration
Simulation reference manipulator and FAM policy configuration.
Sample Size
Test Date
Limitations
Results apply to the stated configuration and defined test conditions. Performance may vary with task complexity, hardware, environment, software version and operator procedures.

Validation details

97% · Defined Base-Task Success Rate in Complex Real Environments

Test Conditions
Defined base tasks in selected real-world environments.
Dataset
Internal real-robot evaluation set.
Task Definition
Complete a predefined task sequence without unrecovered failure.
Evaluation Method
Successful runs divided by total evaluated runs.
Robot Configuration
Anigon FA-L with validated end-effector and model version.
Sample Size
Test Date
Limitations
Results apply to the stated configuration and defined test conditions. Performance may vary with task complexity, hardware, environment, software version and operator procedures.

Validation details

3–5 · Real Demonstrations for Typical Task Adaptation

Test Conditions
Selected typical tasks with stable fixtures and calibrated sensing.
Dataset
Task-specific human demonstration set.
Task Definition
Adapt a pretrained policy to a defined task variant.
Evaluation Method
Count of real demonstrations used before the configured acceptance test.
Robot Configuration
FAM-compatible robot platform and task-specific end-effector.
Sample Size
Test Date
Limitations
Demonstration requirements increase with task complexity, environmental variability and hardware changes.

Validation details

1% · Data Requirement Compared with Selected Traditional Pipelines

Test Conditions
Comparative experiments on selected task pipelines.
Dataset
Internal comparison dataset.
Task Definition
Reach an equivalent configured acceptance threshold.
Evaluation Method
Relative volume of labeled real-world data used by each pipeline.
Robot Configuration
Matched task setup where practical.
Sample Size
Test Date
Limitations
The comparison does not represent every traditional pipeline or every task category.

Validation details

28 DoF · Full-Body Degrees of Freedom

Test Conditions
Product configuration record.
Dataset
Not applicable.
Task Definition
Mechanical degree-of-freedom count for the stated platform configuration.
Evaluation Method
Engineering configuration review.
Robot Configuration
Anigon FA-L reference configuration.
Sample Size
One product configuration.
Test Date
Current configurable specification
Limitations
Final configuration may vary by end-effector, base and customer integration package.

Validation details

24/7 · Continuous Operation with Dual Hot-Swap Batteries

Test Conditions
Operation plan using charged battery rotation and trained service procedures.
Dataset
Product runtime and battery exchange records.
Task Definition
Maintain operational availability through scheduled battery swaps.
Evaluation Method
Availability assessment under the configured duty cycle.
Robot Configuration
FA-L with dual hot-swap battery system.
Sample Size
Deployment dependent.
Test Date
Pilot deployment
Limitations
Actual uptime depends on duty cycle, charging capacity, maintenance, temperature and task load.

Deploy in the real world

Deploy Embodied Intelligence in Your Real-World Operations

From task assessment and data collection to model adaptation, robot deployment, system integration and continuous optimization.

Book a Demo Talk to a Solutions Expert Download Technical Overview