If you've been following humanoid robotics research, you've probably noticed a pattern. Impressive demo videos show robots walking, dancing, or manipulating objects in controlled settings. The papers report high success rates. And then... nothing happens in the real world.
The gap between "works in the lab" and "works in deployment" is, to be precise, a constraints problem. Real environments have obstacles that move. Operators make mistakes. Payloads swing unpredictably. Joint limits exist for reasons. And yet, most control frameworks treat these as afterthoughts, if they address them at all.
Three papers appeared on arXiv this week that tackle different facets of this problem. None of them are revolutionary breakthroughs (I know I'm being picky here, but that term gets overused). What they represent is the kind of incremental, methodical work that actually moves the field toward deployment. Let me walk through what each contributes and where the open questions remain.
The first paper, "A study on a Real-Time VR-Based Teleoperation Framework for Manipulator in Dynamic Environment," addresses something that sounds obvious but apparently isn't: VR teleoperation systems need to handle collisions with moving obstacles.
The authors make a fair point in their framing. Much of the recent VR teleoperation work has focused on data collection for imitation learning. You strap on a headset, move a robot around, and generate demonstrations. The robot doesn't need to be particularly safe during this process because the whole point is just to gather training data. But if you want teleoperation for actual deployment (think hazardous environments, remote operations), you need the system to prevent collisions in real-time, even when the operator makes mistakes.
À lire aussi
More in Humanoids
A wave of new research suggests we can train humanoid robots without expensive human demos. I'm not sure we've thought through what that means.
Sarah Williams · 3 hours ago · 4 min
Two new research papers tackle the same problem from wildly different angles, and honestly, both approaches make me rethink what 'dexterous' really means.
Sarah Williams · 5 hours ago · 6 min
New benchmarks reveal that up to 56% of 'successful' robot manipulation tasks involve safety violations we weren't even tracking.
Sarah Williams · 5 hours ago · 4 min
After years of watching robots stumble because their eyes couldn't keep up with their legs, the research community is finally cracking the perception problem.
Their approach integrates GPU-accelerated inverse kinematics and trajectory optimization directly into the VR control loop. At each control cycle, the system generates feasible joint commands that respect robot constraints while trying to follow the operator's intent. When an obstacle interferes with the commanded path, the system generates detour motions automatically.
The experiments use a 7-DoF manipulator across three scenarios: obstacle-free, static obstacles, and moving obstacles. The results show stable behavior and collision avoidance across all three. It's worth noting that the paper doesn't provide extensive quantitative comparisons to prior methods, which makes it harder to assess exactly how much improvement this represents. The contribution seems to be more architectural (how to structure the real-time optimization pipeline) than algorithmic.
What I'd want to see next: testing with inexperienced operators who actively try to crash the robot. The paper mentions robustness to operator mistakes, but the experimental validation of this claim remains unclear from the abstract.
Here's the issue. Reinforcement learning has produced impressive results for humanoid locomotion and manipulation. But RL policies are, fundamentally, black boxes that map observations to actions. If you train a policy and then realize you need it to avoid a new obstacle, or respect a tighter joint limit, or maintain center of mass stability in a different way, you're stuck. You'd have to retrain the entire policy, which is expensive and might break other behaviors.
The authors present ConstrainedMimic, a framework that enforces constraints at runtime rather than during training. They combine principles from operational space control with control barrier functions (CBFs) to modify the policy's outputs in real-time while staying consistent with the current contact mode and tracking objectives.
Actually, the research shows something subtle here that's easy to miss. The framework doesn't just clamp outputs or apply simple filters. It reasons about whole-body kinematics and dynamics to find the minimal modification that satisfies constraints while preserving as much of the original policy behavior as possible. This is genuinely new compared to simpler constraint enforcement approaches.
The experiments run on a simulated Unitree G1, demonstrating collision avoidance (both self-collision and external obstacles), joint limit enforcement, and center of mass stability constraints during whole-body motion tracking and teleoperation. The system runs at 300-500 Hz on CPU, GPU, or TPU.
A few methodology concerns worth noting. First, all experiments are in simulation. The authors promise to release software, but real-world validation hasn't been replicated yet. Second, the paper focuses on a specific class of constraints (those expressible as CBFs), which may not cover all safety requirements in practice. Third, the "minimal restriction" claim (that the framework doesn't unnecessarily limit policy capabilities) would benefit from more rigorous quantification.
Still, the core idea of post-hoc constraint enforcement for learned policies is valuable. This is the kind of infrastructure work that could make RL-based humanoid control more deployable.
Suspended payloads are underactuated and oscillatory. The robot can only influence the load through whole-body motion and intermittent contact, not through direct actuation. This makes the control problem significantly harder than rigid-body manipulation. Small errors compound. Timing matters. And the dynamics are, in a way, adversarial to naive control strategies.
The HOIST approach combines several pieces. First, they fine-tune a vision-language-action (VLA) policy using VR teleoperation demonstrations. This gives them safe initial behavior that roughly accomplishes the task. Then they use the VLA rollouts combined with iterative batched RL to improve placement accuracy and stopping behavior.
The key insight is that pure imitation learning can get you close but doesn't directly optimize for the final objective (accurate placement). Pure RL from scratch is unsafe and sample-inefficient on real hardware. The hybrid approach uses imitation to bootstrap and RL to refine.
The quantitative results are specific, which I appreciate: compared to pure VLA rollouts, HOIST reduces translational placement error by 19.9 cm and angular error by 3.56 degrees. These aren't small improvements for a task where oscillations make precision difficult.
The experiments include both simulation and real humanoid validation, which is notable. Many papers in this space stay purely in simulation. The real-world results suggest the approach transfers, though the paper doesn't provide detailed analysis of the sim-to-real gap.
One limitation I'd flag: the sample size for real-world experiments is typically small in humanoid papers (hardware time is expensive), and this one appears to follow that pattern. We don't know yet how robust these results are across different payload masses, cable lengths, or environmental conditions.
Looking across all three, I see a common theme: the field is moving from "can we make robots do impressive things?" to "can we make robots do things safely and reliably?"
This is incremental over prior work in the sense that all the underlying techniques (VR teleoperation, CBFs, VLA policies, RL fine-tuning) existed before. What's new is the focus on deployment constraints. Collision avoidance during teleoperation. Post-hoc constraint enforcement for learned policies. Sample-efficient refinement on real hardware.
None of these papers solve the full deployment problem. The teleoperation framework hasn't been tested with truly adversarial operator behavior. The constraint enforcement approach remains simulation-only. The suspended load manipulation results come from a limited number of trials. But collectively, they represent the kind of progress that actually matters for getting humanoids out of the lab.
First, how do these approaches compose? If you have a VLA policy controlling a humanoid that's teleoperating a manipulator while avoiding obstacles and respecting joint limits, do all these constraint enforcement mechanisms play nicely together? The papers address isolated pieces of the problem. Integration remains an open question.
Second, how do they fail? All three papers report success cases. Real deployment requires understanding failure modes. What happens when the constraint satisfaction problem is infeasible? When the RL fine-tuning diverges? When the VR operator's intent conflicts with safety requirements?
Third, what's the computational overhead? The papers report control frequencies (300-500 Hz for ConstrainedMimic, real-time for the teleoperation framework), but the hardware requirements and power consumption for these approaches aren't always clear. For mobile humanoids, this matters.
It's too early to say whether these specific approaches will become standard. But the problems they're addressing (runtime constraints, safe teleoperation, sample-efficient refinement) are the right problems. The field has enough impressive demos. What it needs now is the boring, essential work of making those demos deployable.
(I realize I've written 1,400 words about three papers that might seem narrow. But this is how progress actually happens in robotics. Not through paradigm shifts, but through methodical solutions to specific constraints. The papers that matter most are often the ones that make future papers possible.)