Weights or Skills? A Survey of Robot-Learning Techniques: from Action-Predicting Weights to Robots that Write their Own Skills
Gaytri Jena, Kapil Wanaskar, Vinija Jain, Aman Chadha, Vasu Sharma, Amitava Das
cs.RO, cs.AI
2026-08-03
A survey reorganises robot learning around weights versus skills, and ranks code-as-policy methods by self-improvement (feedback, memory, search); only three systems combine them.
Robot learning has two camps fighting for the steering wheel. One is vision-language-action (VLA) models that press perception, language, and motor control into a single network's weights and trade data and parameters for generalisation, with RT-2, π0, and Gemini Robotics as flagships. The other has large models write executable code on the spot to drive the robot, code that can be revised at run time and banked into a skill library, with Code-as-Policies and Voyager as flagships. The two rarely name the fault line between them, and "making a robot improve with experience" is talked up everywhere but measured nowhere on a shared scale. This survey, submitted to ACM Computing Surveys, draws the line: classify the field by whether competence is baked into weights or written as code, then give the code camp a self-improvement ladder of its own.
From a January 2016 to July 2026 window the authors select 77 systems for the taxonomy plus 225 background "landscape" works, 302 systems in total (310 references including seven prior surveys and one industry source). The 77 are not a grab bag: each must be placeable on the weights/skills axis, contribute a mechanism no other system in the map already carries, and be reported fully enough to fill every comparison column. Systems that fail are demoted to the landscape appendix.
The skeleton is six technique families: code-as-policy, end-to-end VLA, reward synthesis, skill libraries, sim-to-real transfer, and benchmarks. The code camp then gets a self-improvement ladder, the survey's central contribution, built on three mechanisms each given an operational definition that kills boundary disputes:
With these three switches the code camp is sorted into five rungs, thinning sharply toward the top.
The ladder is a clean funnel (Table 4, 30 systems in all):
| Rung | Mechanisms | Systems | Representative |
| Zero-shot synthesis | none of F/M/S | 13 | Code-as-Policies, VoxPoser |
| Closed-loop self-repair | F only | 6 | Inner Monologue, REFLECT |
| Skill-library accumulation | M | 5 | Voyager, Uni-Skill |
| Evolutionary program search | F + S | 3 | CaP-X, RoboEvolve |
| Full self-improving loop | F + M + S | 3 | ASPIRE, ENPIRE, RoboClaw |
The top cell, where feedback, memory, and search combine into one open-ended loop, has only three occupants as of 2026: ASPIRE, ENPIRE, and RoboClaw, all released in 2026. The structural point matters more than any single system: the cell where competence can compound with experience is the emptiest and least-studied region on the map. The VLA camp (11 systems) is marked non-self-improving across the board; competence is frozen at training, generalises by interpolation, and cannot be edited, audited, or recombined afterwards.
"Skill" is badly overloaded, and the survey separates five senses (Table 5): a latent policy (DIAYN), an option or primitive (SPiRL), readable code (Code-as-Policies), a one-tap robot app, and a listed marketplace product. Only the code sense is inspectable, adaptable, and composable at once; only the app and product senses are easy to distribute. What self-improves and what ships are misaligned.
A few numbers worth keeping: Eureka has an LLM write reward functions zero-shot and beats expert-written rewards on 83% of a 29-task IsaacGym suite; Open X-Embodiment pools 60 datasets from 21 institutions into over a million trajectories across 22 embodiments; CrossFormer trains one transformer on 900K trajectories across 20 embodiments and matches per-robot specialists; RoboCat adapts to a new task from as few as 100 demonstrations. Yet every benchmark reports a one-shot success rate, and none plots success as a function of accumulated interaction, which is exactly the quantity the ladder is built to raise.
This is a survey and a position paper, not a new method, and its value is in naming the confusion. For anyone in robotics, "weights versus skills" is now a citable dividing line, and the ladder turns a fuzzy slogan into three checkable switches: when a system claims to self-improve, you can ask which rung it sits on. The sharper move is connecting the taxonomy to a commercial reality already underway. Unitree's UniStore ships one-tap, cross-embodiment motion packages advertised to run on the G1 and H1 humanoids and the B2/Go2 quadrupeds. But every listed skill is static playback, installed and replayed, blind to whether this kitchen, gripper, or object matches the one it was recorded for. The feedback, memory, and search mechanisms on the ladder are precisely what would let a downloaded skill adapt on site, so today's store is a library of animations, not capabilities. The authors name this static-to-adaptive gap as the defining research opportunity for physical AI.
The authors state their boundaries plainly. First, the self-improvement axis is deliberately code-centric; VLA models and RL skill discovery are treated as context rather than protagonists, and a survey centred on representation learning or data collection would draw a different map. Second, the three top-rung systems are 2026, some concurrent, with non-comparable benchmarks, so the survey reports capabilities qualitatively and refuses to rank. Third, the "skill economy" framing rests on a single vendor's marketplace and is motivation, not mature results.
Two more doubts from the reading. The 77-system selection relies on human judgement of representativeness and analytical role, and rules like "taxonomise only the earliest landmark making a given point" invite selection bias, so the funnel counts reflect coverage rather than the field's true proportions; the authors say so, but readers may read it as a league table. And "the code camp is the real self-improvement path" is a stance, not a fact: the survey concedes the transfer and generalisation strengths of VLAs (million-trajectory, cross-embodiment scale), and labelling that whole camp "non-self-improving" measures it with the code camp's own ruler. The three top-rung systems are too new to be battle-tested.