🤖 Hardware: Cobots, Humanoids, and Platforms
From the collaborative arms that dominate research datasets to cutting-edge humanoids — and the accessible platforms that let you practice VLA for real. The hardware you choose determines the action space, the available sensors, and the compute required.
Illustrative diagram — cobots, humanoids, platforms, and sensors all feed into a VLA policy.
🦾 Industrial Cobots: Franka, UR, KUKA
Cobots (collaborative robots) are robotic arms designed to operate safely alongside humans, without safety cages. The Franka Emika (7 DoF, torque-sensitive), Universal Robots (UR3/UR5/UR10), and the KUKA LBR iiwa dominate VLA research labs and are the basis for most large datasets.
🔑 Why cobots dominate VLA research
- Franka7 DoF with torque sensing at each joint. Avoids singularities and enables force control. Dominates the DROID dataset with over 76,000 episodes.
- URUniversal Robots UR3/5/10 — simple, affordable, and widely used in light industrial automation and data collection setups.
- KUKALBR iiwa with 7 axes and force/torque sensing on every axis. A reference for impedance control research.
✓ Advantages of cobots in VLA
- ✓Open API (ROS, libfranka) makes integration with models easier
- ✓Torque sensors enable contact detection and precise control
- ✓Huge datasets with demonstrations ready for fine-tuning
✗ Limitations to know about
- ✗No locomotion—fixed to a base, they don't move around the environment
- ✗Workspace limited to arm’s reach (~800 mm radius)
- ✗Cost of $10k–$50k — inaccessible without institutional support
Key concepts
They define mobility. 6 DoF for reaching an arbitrary pose; 7 DoF add redundancy and avoid singularities.
A sensor at each joint measures the force applied—essential for impedance control and collaborative safety.
Low-level interfaces that connect the AI model to the arm’s real-time controller.
🧍 Humanoids: Figure, Unitree, Optimus, 1X
Bipedal humanoid robots designed to operate in environments built for humans. Figure 02, Unitree G1/H1, Tesla Optimus and the robots from 1X combine bipedal locomotion, dexterous manipulation, and increasingly, control by VLAs. They are the most ambitious and visible frontier of modern robotics.
tree
Unitree G1/H1 — democratizing humanoid robots
The G1 costs ~$16k, lowering the barrier to entry for research labs. With 23 DoF and stable bipedal locomotion, it’s currently the most accessible humanoid for anyone looking to experiment with VLA control.
Figure 02 — arms in BMW factories
Already operates in a real industrial environment. Uses VLA to interpret instructions and manipulate parts. Partnership with OpenAI for the language module. A reference for deploying VLA in production.
Tesla Optimus — scale as leverage
Tesla plans mass production, using the same data collection pipeline as its self-driving cars to train Optimus. Scaling demonstration data is the central bet.
1X NEO — focus on whole-body fluency
1X bet on robots with a more organic form (wheels + arms) before moving to bipedal robots. NEO integrates end-to-end VLA, focusing on smooth, safe movements around people.
💡 Calibrating expectations
Humanoids make impressive videos, but the whole-body control — coordinate locomotion and manipulation simultaneously — still very difficult. Most current VLAs control stationary arms. Bipedal locomotion on real robots is usually delegated to separate controllers.
✋ Grippers and end effectors
O end-effector is the robot’s final interface with the physical world—its “hand.” It ranges from simple parallel grippers (two jaws) to dexterous anthropomorphic hands with dozens of DoF. The choice of end effector directly determines what the VLA needs to control as output and which tasks are physically feasible.
✓ Parallel gripper — simple and reliable
- ✓1 DoF (open/close) — minimum action space in the VLA
- ✓Robust, inexpensive, easy to calibrate and replace
- ✓Sufficient for pick-and-place, the most common task in datasets
✗ Simple gripper limitations
- ✗Cannot manipulate deformable objects (fabric, wires)
- ✗No control over finger orientation—limited grasping poses
- ✗Dexterous tasks (screws, keyboards) require complex hands
📊 End-effector spectrum by complexity
Robotiq 2F-85. Standard on Franka and UR. Action space: 1 number between 0 and 1.
Robotiq 3-Finger, Allegro Hand. Enables object reorientation. Much harder to control.
Shadow Hand. Artificial human hand. Extremely complex — cutting-edge research.
💡 The end-effector defines the action space
The VLA's output is an action vector. A parallel gripper adds 1 dimension; a dexterous hand adds 20+. The more DoF the end-effector has, the more difficult the learning problem. To get started, use a simple gripper and tackle complex tasks with an intelligent policy.
👁️ Sensors: RGB-D, force/torque, tactile
The robot's senses form the perceptual input to the VLA. RGB-D cameras provide color and depth; wrist force/torque sensors measure interaction with objects; tactile sensors capture texture and contact at the fingertips. The combination of these modalities—and what you leave out—directly affects what the policy can learn.
Illustrative diagram — most current VLAs use only RGB; force and touch are research frontiers.
Key concepts by modality
The depth (D) channel adds 3D pose estimation without extra physical structure. Intel RealSense D435 is the standard for manipulation. Limitation: occlusion and reflective surfaces.
A sensor at the wrist measures applied forces and moments. It enables impedance control (yielding to force) and contact detection—essential for part insertion and delicate manipulation.
GelSight and DIGIT capture images of contact at the fingertips, revealing texture and slip. Still not widely integrated into mainstream VLAs, but an active area of multimodal research.
🔓 Affordable platforms: SO-100, Koch, ALOHA
Low-cost robotic arms low cost and open source created to democratize VLA learning. The SO-100/SO-101 (Hugging Face’s LeRobot project) costs a few hundred dollars; Koch and ALOHA use leader-follower setups for accessible teleoperation and data collection. They are a realistic way to practice VLA with physical hardware.
🔑 The SO-100 + LeRobot stack
- SO-1003D-printed 6 DoF arm with affordable servomotors. Total cost <$500. The leader arm enables intuitive teleoperation for collecting demonstrations.
- LeRobotHugging Face library that brings together hardware, datasets, and trainable policies (ACT, Diffusion Policy, SmolVLA) in a single pipeline. Deploying a policy trained in 1 day is feasible.
- ALOHALow-cost bimanual setup (two arms) developed at Stanford. Used to train ACT and Mobile ALOHA. A reference for dexterous manipulation on a limited budget.
✓ Why start here
- ✓Real physical hardware — learns problems that simulation hides
- ✓In-house data collection — custom datasets for your task
- ✓Active LeRobot community on Discord with support and examples
✗ What to expect (realism)
- ✗Lower mechanical precision than industrial cobots—requires frequent calibration
- ✗Low payload (<500 g) — only for lightweight benchtop objects
- ✗No torque sensors—much more limited force control
💡 The leader-follower configuration
In a leader-follower setup, you move a twin arm by hand and the robot copies the movement in real time, recording each frame as a demonstration. This approach captures high-quality data without programming skills, making it feasible to collect 50–100 episodes over a weekend.
⚡ Embedded Compute: Jetson and Latency
The compute hardware that runs the VLA, often embedded in the robot itself. The NVIDIA Jetson line (Orin Nano, AGX Orin) is the standard for on-device inference: a GPU in a compact form factor, low power consumption, viable for autonomous deployment without an external server. The choice of compute determines whether the VLA runs in real time or needs a remote data center.
Comparison: edge vs. cloud
- • Latency <20 ms — compatible with 50 Hz control
- • Autonomous — no network dependency
- • AGX Orin: 275 TOPS at 60W
- • TensorRT + INT8 quantization required
- • Larger models (7B+) with no memory constraints
- • Latency of 50–200 ms over Wi-Fi — limit of ~10 Hz
- • Fails if the network goes down — a risk in real-world environments
- • Good for prototyping, poor for final deployment
✓ Edge optimization techniques
- ✓INT8/INT4 quantization — reduces memory use and speeds up inference with little loss
- ✓TensorRT — compiles the model for the Jetson GPU with specific optimizations
- ✓Smaller models (SmolVLA, 450M) run comfortably on the Orin
✗ Common pitfalls
- ✗Test only on a research GPU and assume it runs on the Jetson — it doesn’t
- ✗Ignore camera latency — frame buffers add an extra 30–80 ms
- ✗Float32 model on Jetson without quantization — insufficient VRAM above 3B
Key concepts
Frequency at which the robot reads sensors and sends new commands. 10 Hz = 100 ms maximum end-to-end pipeline latency.
Reduces the numerical precision of weights (float32 → INT8). Compresses the model 4× and speeds it up 2–4× on dedicated hardware.
Pits autonomy and guaranteed on-device latency against greater capability with the risk of delay and network loss.
✅ Module Summary
Next module
1.6 — Safety and Standards: ISO 10218:2025