ExplorRobo
robot-manipulationaction-spaceforce-sensingworld-action-modelsrobot-learningphysical-aihumanoids

End Effector Poses Buy Generality, Force Buys Throughput

By ExplorRobo Editorial · · 11 min read

Figures attributed to companies, projects and papers are their published claims. The synthesis and judgements below are this author’s analysis. I read the announcements, research notes and paper abstracts; I have not run any of these systems.

On 20 October, Eka Robotics announced a 150 million dollar Series A for a vision-force-action model, arguing that performant manipulation comes from simulation and reinforcement learning rather than human demonstration data. Days later, Dream Labs came out of stealth with 90 million dollars and Dream-1, a model that controls bimanual robots it has never seen, including a PR2 built 18 years ago, from three camera views and a prompt.

Both call the result general. The difference is not model size or data volume, but what a policy is allowed to output. Eka’s action includes force. Dream-1’s does not. That choice explains why Dream-1 transfers across two dozen arm models while Eka claims nothing of the sort, why one sells an API at three dollars an hour and the other sells deployments, and why their headline numbers measure different things. One buys contact authority. The other buys embodiment freedom.

A pose, not a wrench

Dream Labs describes the mechanism plainly. A policy usually outputs joint angles, which ties it to one body. Dream-1 does that for the robots it supports natively, but it can also predict end effector poses: where each gripper should be, and whether it should be open or closed. Every robot with two arms and parallel grippers shares that space, and its own inverse kinematics solver turns poses into joint commands.

The shared interface is necessary but not sufficient. Transfer works because Dream-1 sees through three cameras, a head view and one wrist view per arm, and its world action model predicts video and actions jointly, so every robot looks the same to the head camera: pixels. The precedent is the Universal Manipulation Interface paper from RSS 2024, where policies transferred across arms that shared one gripper and relied on wrist cameras only. Dream-1 removes that constraint by treating the body as video.

Dream Labs lists the costs: IK quality, an arm path the model does not choose, and infeasible predictions on unfamiliar kinematics. There is a fourth cost it does not list.

An end effector pose plus a binary open or closed command has no channel for force. Not as an input, which is a separate question, but as an output. The policy can say where to put the gripper and when to close it. It cannot say how hard to press, how stiff the wrist should be, or what to do when the wrench at the fingers says the object is slipping. A robot running Dream-1 can carry a wrist force and torque sensor, but the policy cannot request compliance or react to a contact it did not plan. Compliance becomes a property of the hardware, fixed before the model runs.

That is the trade that makes the product possible, and it recurs wherever a policy must be body agnostic.

Eka built the hand it needed

Eka’s argument is that performance comes “not by mimicking human data, but by mastering physics”, with AlphaZero as the precedent, because it used zero human data. Simulation plus reinforcement learning, it says, generates “thousands of hours of robot experience in minutes”, and the sim-to-real gap for contact-rich work is closing.

The published result is throughput: 1,500 pick operations per hour against a human baseline of roughly 600, on objects outside the training distribution, and the same robot can screw in a light bulb and insert wires. TechCrunch adds what the post omits: a 1.45 billion dollar post-money valuation, objects up to one kilogram, a success claim above 95 percent, and a custom hand with more force sensors than a standard gripper, which the founders would not detail.

Read the generality claim closely. Eka’s site promises generalization “across objects, tasks, and environments”. Robots are not on that list. Dream-1’s note promises unseen environments, unseen tasks “and even robots”. One adjective, two axes.

A policy whose action includes finger level force control is bound to the hand that produced the signal, its sensor layout, its compliance and its controller. The hand is part of the action space, which is why Eka had to build one: a force channel cannot be bolted onto an existing gripper, because what transfers includes the mechanical interface. Eka buys contact authority and pays in embodiment lock-in. Dream Labs buys the reverse.

The one number measured outside both companies

Eka’s figures are company numbers on the company’s own setup, with no task list, evaluation protocol or failure taxonomy in the launch materials. That does not make them wrong. It makes them uncheckable.

Dream-1 has a number that is not Dream Labs’ own. Robocurve, a public benefit corporation that publishes frontier robotics evaluations, reports Dream-1 at a mean score of 59.3 percent on its RoboDojo-RC Tier 1 suite, at 0.15 dollars and 1.5 minutes per task attempt, over 120 attempts. The next best score is Gemini 4 Argon at 48.2 percent, at 2.34 dollars and 20.5 minutes per attempt. GR00T N1.7, an actual vision-language-action policy, scores 9.5 percent.

Three things follow. The cost and latency gap is not a rounding error: the model returns 1.6 seconds of actions per call, priced near 0.0013 dollars per action. Second, 59.3 percent is the honest state of the field on sorting, stacking, pouring and storing tasks, so any claim of solved manipulation should be read against a number in the fifties. Third, Robocurve separates its runs by control mode, end effector or joint position, and its six Tier 1 tasks are pick, place, sort and pour work. Nothing there turns on a force channel, so the benchmark measures the layer Dream-1 optimizes. That is fair, and it is also why the architecture comparison remains one measured number against one unmeasured claim.

Force in code, for everyone else

You do not need a force-native foundation model to get contact behaviour. FORCA, posted to arXiv on 22 October by four researchers at the University of Illinois Urbana-Champaign, is a coding agent that takes visual observations and wrist force and torque readings and writes Python programs for contact probing and force-responsive motion. During a rollout the program uses the wrench to regulate motion; between rollouts the agent compares force traces with the visual outcome and rewrites both the motion and the force feedback logic.

It needs no demonstrations, no task specific training and no simulation. On five real robot tasks (drawer opening, door handle operation, cube knocking, vase wiping, cucumber peeling) it beat a vision-only variant of the same agent on every dimension the authors report: drawer opening success rose from 40 to 100 percent, door handle operation from 10 to 100 percent, and contact loss in wiping and peeling fell from 40.6 to 55.9 percent to at most 2.2 percent.

The ablation compares two configurations of one agent, so read this as a mechanism demonstration with a striking magnitude: the only variable that changed was whether the program could read force.

This is the split argued in The Model Proposes, the Gate Disposes, applied to contact. Force feedback lives in code that can be logged, diffed and versioned rather than in weights nobody can audit, and a developer with an ordinary arm and a wrist sensor can use it without solving cross-embodiment transfer first.

Agility’s Digit 5 is the reminder that the physical terms still dominate. Its published specification is 23 kilograms of payload, a 10 to 1 run to charge ratio and 20 working hours in a 24 hour day, on a robot the company says was trained on 65,000 hours of operational data, and its co-founder wrote in IEEE Spectrum that the shipped machine is “heavy and bulky” compared with the one he dreamed of.

What this means if you are building

  • Ask for the action space before the model. A pose plus open or closed command means compliance is your controller’s job, permanently. Pick an arm with a documented Cartesian impedance mode, and know the policy will not modulate it.
  • Log the wrench even if nothing consumes it. A wrist force and torque sensor costs less than doing the integration twice, and those logs are what a FORCA style loop will need.
  • Test transfer claims on your gripper. Dream-1’s supported robots are mostly parallel grippers. If your end effector differs, read the fine-tuning numbers first: 5 minutes of data for a YAM and a Maker Arm, 77 minutes for Dexmate with Sharpa hands, 98 minutes for a YAM with unusual end effectors.

What to watch

Whether Eka publishes an evaluation protocol, a task list or a paper, because a 1.45 billion dollar valuation resting on one throughput figure is a lot of trust in an unaudited number.

Whether Dream-1’s action space grows a compliance channel. The obvious version is not exotic: add a desired wrench, or a Cartesian stiffness and damping vector, to the pose command. Impedance controlled arms already accept commands in that form, it stays body agnostic, and a policy could then ask for soft contact. The missing piece is a convention, not a breakthrough, and a convention is what UMI provided for grippers.

Whether anyone runs a force-native and a pose-native policy on the same contact-rich suite. Robocurve publishes its runs and separates control modes, but its tasks would need at least one peeling, wiping or insertion task before the comparison means anything.

And whether the open data layer catches up. NVIDIA’s FLORA, released the same week, labels 200,000 hours per day across 512 GB300s and claims to be roughly 125 times cheaper than GPT-6 Astra. Its outputs are captions and reward labels, the ingredients a video-trained policy needs. There is no equivalent open pipeline for force, because a wrench log means nothing without the hand and controller that produced it, so the video-first camp keeps scaling data faster than the force-first camp can scale sensors. Your own acceptance test still decides, which was the argument in The Robot Data Refinery Is the Real Product.

Sources

Frequently asked questions

What is an end effector pose action space?
Instead of predicting joint angles, the policy outputs where each gripper should be in space and whether it should be open or closed. The robot's own inverse kinematics solver turns that pose into joint commands. Dream Labs documents this as the mechanism behind Dream-1's cross-embodiment transfer, and lists solver quality as one of its limits.
Can a vision-only policy learn contact-rich tasks?
It can attempt them, but it has no channel to regulate contact. The FORCA paper from UIUC measured the gap on five real robot tasks: adding wrist force and torque input raised drawer opening success from 40 to 100 percent and door handle operation from 10 to 100 percent against a vision-only variant of the same agent.

Related reading

World-Action Models Agree on the Bottleneck, Not the Fix

A vendor launch and four arXiv papers attacked the robot data bottleneck in one week, and disagree about where a policy's generality comes from.

Robot imports become a security question

Washington opens trade talks on imported robots, Agility starts trading as AGLT, and two open-weight models rebuild manipulation data from force and gloves.

The Robot Stack Is Going Modular, and the Guard Layer Has No Home

HomeBody deletes the learned policy, FLUX 3 Action shrinks it to a checkpoint, and RoboHarm shows the frontier model that takes over refuses almost nothing.